@prefix skos: . @prefix dct: . @prefix owl: . @prefix mlsys: . a skos:ConceptScheme ; dct:title "ML Systems Taxonomy"@en ; owl:versionInfo "0.2-draft" ; dct:creator "Vijay Janapa Reddi"@en ; dct:license ; dct:description "Editorial draft for expert review; not a claim of completeness or consensus."@en . mlsys:specializes a owl:ObjectProperty ; owl:subPropertyOf skos:related ; skos:prefLabel "is a kind of"@en ; skos:definition "A is a narrower or more specific case of B."@en . mlsys:part-of a owl:ObjectProperty ; owl:subPropertyOf skos:related ; skos:prefLabel "is part of"@en ; skos:definition "A is a component or stage within B."@en . mlsys:alternative-to a owl:ObjectProperty ; owl:subPropertyOf skos:related ; skos:prefLabel "is an alternative to"@en ; skos:definition "A and B are competing options for the same job (symmetric)."@en . mlsys:requires a owl:ObjectProperty ; owl:subPropertyOf skos:related ; skos:prefLabel "requires"@en ; skos:definition "A cannot work, or cannot be correct, without B."@en . mlsys:uses a owl:ObjectProperty ; owl:subPropertyOf skos:related ; skos:prefLabel "uses"@en ; skos:definition "A commonly relies on B as a building block, without strict necessity."@en . mlsys:implements a owl:ObjectProperty ; owl:subPropertyOf skos:related ; skos:prefLabel "implements"@en ; skos:definition "A is a concrete mechanism or technique that realizes B."@en . mlsys:enables a owl:ObjectProperty ; owl:subPropertyOf skos:related ; skos:prefLabel "enables"@en ; skos:definition "A makes B possible or practical."@en . mlsys:produces a owl:ObjectProperty ; owl:subPropertyOf skos:related ; skos:prefLabel "produces"@en ; skos:definition "A creates or emits B as an output or artifact."@en . mlsys:reduces a owl:ObjectProperty ; owl:subPropertyOf skos:related ; skos:prefLabel "reduces"@en ; skos:definition "A lowers the amount, cost, or magnitude of B."@en . mlsys:increases a owl:ObjectProperty ; owl:subPropertyOf skos:related ; skos:prefLabel "increases"@en ; skos:definition "A raises the amount, cost, or magnitude of B."@en . mlsys:determines a owl:ObjectProperty ; owl:subPropertyOf skos:related ; skos:prefLabel "determines"@en ; skos:definition "A sets or shapes B (direction not simply up/down)."@en . mlsys:limits a owl:ObjectProperty ; owl:subPropertyOf skos:related ; skos:prefLabel "limits"@en ; skos:definition "A places a limit or ceiling on B."@en . mlsys:trades-off a owl:ObjectProperty ; owl:subPropertyOf skos:related ; skos:prefLabel "trades off against"@en ; skos:definition "Improving A tends to worsen B, or vice versa (symmetric)."@en . mlsys:measures a owl:ObjectProperty ; owl:subPropertyOf skos:related ; skos:prefLabel "measures"@en ; skos:definition "A quantifies, evaluates, or tests B."@en . mlsys:detects a owl:ObjectProperty ; owl:subPropertyOf skos:related ; skos:prefLabel "detects"@en ; skos:definition "A reveals or identifies occurrences of B."@en . mlsys:causes a owl:ObjectProperty ; owl:subPropertyOf skos:related ; skos:prefLabel "can cause"@en ; skos:definition "A can lead to the failure, harm, or degradation B."@en . mlsys:mitigates a owl:ObjectProperty ; owl:subPropertyOf skos:related ; skos:prefLabel "mitigates"@en ; skos:definition "A protects against, prevents, or recovers from B."@en . a skos:Collection ; skos:prefLabel "Foundations"@en ; skos:notation "F" ; skos:member , , . a skos:Collection ; skos:prefLabel "Workloads & requirements"@en ; skos:notation "F1" ; skos:scopeNote "Targets, workload descriptions, and objectives that every later engineering choice is judged against."@en ; skos:member , , . a skos:Collection ; skos:prefLabel "Requirements, SLOs & baselines"@en ; skos:notation "F1.1" ; skos:definition "Stating measurable targets for a system's behavior and service, and the baselines and budgets against which they are judged."@en ; skos:member , , , , , . a skos:Collection ; skos:prefLabel "Workload characterization"@en ; skos:notation "F1.2" ; skos:definition "Describing a workload's compute, memory, data, and traffic demands, and where it runs, so it can be engineered and provisioned for."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Objectives, proxies & decision policies"@en ; skos:notation "F1.3" ; skos:definition "How learned outputs become decisions, how optimized proxies relate to intended outcomes, and how competing objectives are traded."@en ; skos:member , , , . a skos:Collection ; skos:prefLabel "Models & architectures"@en ; skos:notation "F2" ; skos:scopeNote "Model families, building blocks, and architectural choices viewed as workloads, plus the artifacts that carry them."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Model families"@en ; skos:notation "F2.1" ; skos:definition "The classical and neural model regimes, with their inductive assumptions, viewed as workloads with characteristic compute and memory profiles."@en ; skos:member , , , , , , , , , , , , , . a skos:Collection ; skos:prefLabel "Architecture building blocks"@en ; skos:notation "F2.2" ; skos:definition "The layer-level operations from which neural models are composed and whose shapes determine compute, memory, and data movement."@en ; skos:member , , , , , , , , . a skos:Collection ; skos:prefLabel "Sparse & conditional computation"@en ; skos:notation "F2.3" ; skos:definition "Model designs in which the computation performed varies per input, such as mixture-of-experts routing and early exit."@en ; skos:member , , , , , . a skos:Collection ; skos:prefLabel "Long-context & efficient sequence models"@en ; skos:notation "F2.4" ; skos:definition "Sequence architectures that change how cost scales with context length, co-designed with the memory and compute they demand."@en ; skos:member , , , , , , , . a skos:Collection ; skos:prefLabel "Model artifacts & formats"@en ; skos:notation "F2.5" ; skos:definition "The weights, graphs, metadata, and interface contracts that make a trained model a portable, loadable, deployable artifact."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Quantitative system models"@en ; skos:notation "F3" ; skos:scopeNote "Analytical and empirical models used to bound performance, capacity, and training cost before or instead of measuring."@en ; skos:member , , , . a skos:Collection ; skos:prefLabel "Roofline, performance models & simulation"@en ; skos:notation "F3.1" ; skos:definition "Bound-based reasoning about whether computation is limited by arithmetic or data movement, and how reuse changes that limit."@en ; skos:member , , , , , , , , , . a skos:Collection ; skos:prefLabel "Amdahl's law & scaling efficiency"@en ; skos:notation "F3.2" ; skos:definition "Models of how much speedup added parallel or accelerated resources can deliver, given serial work and fixed or growing problem sizes."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Queueing & capacity models"@en ; skos:notation "F3.3" ; skos:definition "Models relating arrivals, service rates, utilization, and waiting, used to predict latency and capacity under load."@en ; skos:member , , , , , , . a skos:Collection ; skos:prefLabel "Scaling laws & compute-optimal allocation"@en ; skos:notation "F3.4" ; skos:definition "Empirical relationships among parameters, data, and compute and observed quality, used to allocate training budgets."@en ; skos:member , , , . a skos:Collection ; skos:prefLabel "Stack"@en ; skos:notation "S" ; skos:member , , , . a skos:Collection ; skos:prefLabel "Hardware & accelerators"@en ; skos:notation "S1" ; skos:scopeNote "Compute, memory, interconnect, and power characteristics of ML hardware from datacenter accelerators to microcontrollers."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Accelerator architecture"@en ; skos:notation "S1.1" ; skos:definition "The organization of processors specialized for ML: execution units, threading, dataflow, and the throughput they provide."@en ; skos:member , , , , , , , , . a skos:Collection ; skos:prefLabel "Memory hierarchy & bandwidth"@en ; skos:notation "S1.2" ; skos:definition "The storage levels of an ML processor, from registers and SRAM to HBM, and the capacity and bandwidth limits they impose."@en ; skos:member , , , , , , , . a skos:Collection ; skos:prefLabel "Host-device & scale-up interconnect"@en ; skos:notation "S1.3" ; skos:definition "Links that move data between hosts and accelerators and among accelerators within a node or rack-scale domain."@en ; skos:member , , , . a skos:Collection ; skos:prefLabel "Chip power & thermal limits"@en ; skos:notation "S1.4" ; skos:definition "Electrical and thermal limits at the chip and board level, and the runtime controls that keep execution within them."@en ; skos:member , , , , , , . a skos:Collection ; skos:prefLabel "Edge & embedded hardware"@en ; skos:notation "S1.5" ; skos:definition "Hardware for on-device ML, from mobile SoCs to microcontrollers, where memory and energy per inference bind."@en ; skos:member , , , , , . a skos:Collection ; skos:prefLabel "Frameworks, compilers & runtimes"@en ; skos:notation "S2" ; skos:scopeNote "The software layers between model code and hardware: programming models, operators, compilers, runtimes, and numerics."@en ; skos:member , , , , , , . a skos:Collection ; skos:prefLabel "ML frameworks & programming models"@en ; skos:notation "S2.1" ; skos:definition "The abstractions with which ML programs are written: tensors, computation graphs, execution modes, and automatic differentiation."@en ; skos:member , , , , , , , , , , , . a skos:Collection ; skos:prefLabel "Tensor operators & memory layout"@en ; skos:notation "S2.2" ; skos:definition "Implementations of core tensor operations, such as matrix multiply, convolution, and attention, and the layouts that govern their efficiency."@en ; skos:member , , , , , , . a skos:Collection ; skos:prefLabel "Kernel & GPU programming model"@en ; skos:notation "S2.3" ; skos:definition "How kernels are written for parallel accelerators: the thread hierarchy, memory access patterns, occupancy, asynchronous data movement, and programming the matrix units."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "ML compilers & kernel optimization"@en ; skos:notation "S2.4" ; skos:definition "Transforming model programs through intermediate representations into optimized kernels for target hardware."@en ; skos:member , , , , , , , , , , . a skos:Collection ; skos:prefLabel "Runtime execution & dispatch"@en ; skos:notation "S2.5" ; skos:definition "How a runtime launches and orders device work and manages device memory during execution."@en ; skos:member , , , , , , , , , . a skos:Collection ; skos:prefLabel "Data loading & input pipelines"@en ; skos:notation "S2.6" ; skos:definition "Delivering training and inference inputs to accelerators fast enough: host-side loading and preprocessing, prefetching, and sampling minibatches from large or irregular data such as graphs."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Numerical formats & precision"@en ; skos:notation "S2.7" ; skos:definition "Number formats used in ML computation and the techniques that keep training and inference accurate at reduced precision."@en ; skos:member , , , , , , , . a skos:Collection ; skos:prefLabel "Parallelism & distributed execution"@en ; skos:notation "S3" ; skos:scopeNote "Partitioning ML work across devices and participants, the communication it requires, and the coordination that keeps it consistent."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Parallelism strategies"@en ; skos:notation "S3.1" ; skos:definition "Ways of partitioning a model's computation and state across devices, and how those dimensions are combined."@en ; skos:member , , , , , , , , , , , . a skos:Collection ; skos:prefLabel "Parallelization planning & placement"@en ; skos:notation "S3.2" ; skos:definition "Choosing and changing how a model is partitioned across devices: describing devices as a mesh, searching or propagating sharding decisions, placing experts and stages on heterogeneous hardware, and resharding at run time."@en ; skos:member , , , , , . a skos:Collection ; skos:prefLabel "Collective communication"@en ; skos:notation "S3.3" ; skos:definition "Group communication primitives and algorithms used by parallel ML, their cost models, and techniques to reduce or hide their cost."@en ; skos:member , , , , , , , , , . a skos:Collection ; skos:prefLabel "Coordination & consensus"@en ; skos:notation "S3.4" ; skos:definition "Mechanisms by which distributed components agree on authoritative control state and tolerate stale or failed participants."@en ; skos:member , , , , , . a skos:Collection ; skos:prefLabel "Federated & decentralized learning"@en ; skos:notation "S3.5" ; skos:definition "Training across devices or organizations that keep their data local, coordinated through a shared model."@en ; skos:member , , , , , . a skos:Collection ; skos:prefLabel "Datacenter infrastructure & fleet"@en ; skos:notation "S4" ; skos:scopeNote "Facilities, networks, storage, and cluster management that turn many accelerators into a shared ML fleet."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Datacenter power, cooling & facilities"@en ; skos:notation "S4.1" ; skos:definition "Engineering of the buildings that house ML fleets: power provisioning, cooling, density, and facility efficiency."@en ; skos:member , , , , , , , , . a skos:Collection ; skos:prefLabel "Network fabrics"@en ; skos:notation "S4.2" ; skos:definition "Datacenter networks that connect accelerator nodes: transports, topologies, flow and congestion control, and in-network compute."@en ; skos:member , , , , , , , , , , . a skos:Collection ; skos:prefLabel "Storage & data delivery"@en ; skos:notation "S4.3" ; skos:definition "Storage systems that hold training data, checkpoints, and artifacts at scale and deliver them to accelerators fast enough."@en ; skos:member , , , . a skos:Collection ; skos:prefLabel "Cluster scheduling & orchestration"@en ; skos:notation "S4.4" ; skos:definition "Allocating a shared cluster's accelerators to jobs over time: placement, gang allocation, fairness, preemption, and elasticity."@en ; skos:member , , , , , , , . a skos:Collection ; skos:prefLabel "Capacity planning & multi-tenancy"@en ; skos:notation "S4.5" ; skos:definition "Provisioning fleet capacity and sharing it among tenants through quotas, accelerator sharing, and interference control."@en ; skos:member , , , , , , . a skos:Collection ; skos:prefLabel "Lifecycle"@en ; skos:notation "P" ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Data engineering"@en ; skos:notation "P1" ; skos:scopeNote "Acquisition and labeling, corpus curation, validation and features, pipelines and storage formats, lineage, and temporal correctness of training and serving data."@en ; skos:member , , , , , . a skos:Collection ; skos:prefLabel "Data collection & labeling"@en ; skos:notation "P1.1" ; skos:definition "Obtaining observations and supervisory targets whose sampling, coverage, and label quality are known and fit the task."@en ; skos:member , , , , , , . a skos:Collection ; skos:prefLabel "Data selection & curation"@en ; skos:notation "P1.2" ; skos:definition "Choosing, filtering, deduplicating, mixing, augmenting, and synthesizing training examples to maximize learning value per unit of compute and labeling cost."@en ; skos:member , , , , , , . a skos:Collection ; skos:prefLabel "Data validation & feature engineering"@en ; skos:notation "P1.3" ; skos:definition "Checking that recorded data is fit for use and transforming it into model inputs whose semantics and availability match training and serving."@en ; skos:member , , , , , , , , , . a skos:Collection ; skos:prefLabel "Data pipelines & storage formats"@en ; skos:notation "P1.4" ; skos:definition "Engineering ingestion, batch and stream transformation, event-time semantics, storage formats, and schemas for training and serving data."@en ; skos:member , , , , , , , , . a skos:Collection ; skos:prefLabel "Data contracts, versioning & lineage"@en ; skos:notation "P1.5" ; skos:definition "Specifying, versioning, and tracing the datasets a system consumes so that dependencies, provenance, and changes are explicit."@en ; skos:member , , , , , , , . a skos:Collection ; skos:prefLabel "Leakage & training-serving skew"@en ; skos:notation "P1.6" ; skos:definition "Ensuring examples use only information available at prediction time and that training and serving compute features identically and equally fresh."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Training & adaptation"@en ; skos:notation "P2" ; skos:scopeNote "Learning paradigms, optimization dynamics, training memory, experiment management, fine-tuning, post-training, and continual learning."@en ; skos:member , , , , , , . a skos:Collection ; skos:prefLabel "Learning paradigms"@en ; skos:notation "P2.1" ; skos:definition "The ways supervision, self-generated targets, interaction, and demonstration produce learning signals, each with distinct data and compute demands."@en ; skos:member , , , , , , , , , . a skos:Collection ; skos:prefLabel "Optimization & training dynamics"@en ; skos:notation "P2.2" ; skos:definition "The objectives, update rules, schedules, and batch regimes that determine whether and how efficiently training converges."@en ; skos:member , , , , , , , , , . a skos:Collection ; skos:prefLabel "Training memory & state"@en ; skos:notation "P2.3" ; skos:definition "Accounting for and reducing the parameters, gradients, optimizer state, and activations a training step must hold."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Hyperparameter search & experiment management"@en ; skos:notation "P2.4" ; skos:definition "Allocating a finite budget of training runs across configurations, and recording runs so their outcomes can be compared and reproduced."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Adaptation: fine-tuning, prompting & transfer"@en ; skos:notation "P2.5" ; skos:definition "Adapting a pretrained model to a new task or behavior, by continued training of all or part of its parameters or by prompting without parameter updates."@en ; skos:member , , , , , , . a skos:Collection ; skos:prefLabel "Preference & RL post-training"@en ; skos:notation "P2.6" ; skos:definition "Training pretrained models against preference, reward-model, or verifiable-reward signals, and the rollout, scoring, and weight-synchronization systems that loop requires."@en ; skos:member , , , , , , , , . a skos:Collection ; skos:prefLabel "Continual & online learning"@en ; skos:notation "P2.7" ; skos:definition "Updating deployed models incrementally from new or streaming data while retaining prior competence."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Model evaluation & testing"@en ; skos:notation "P3" ; skos:scopeNote "Evaluation design, metrics and uncertainty, slices and robustness, evaluation of generative and agentic outputs, and ML testing and debugging."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Evaluation design & experiments"@en ; skos:notation "P3.1" ; skos:definition "Constructing offline splits and online experiments whose comparisons support the conclusions drawn from them."@en ; skos:member , , , , , , , . a skos:Collection ; skos:prefLabel "Metrics, calibration & uncertainty"@en ; skos:notation "P3.2" ; skos:definition "Specifying what is measured and quantifying how far estimates and predicted probabilities can be trusted."@en ; skos:member , , , , , , . a skos:Collection ; skos:prefLabel "Sliced & robustness evaluation"@en ; skos:notation "P3.3" ; skos:definition "Measuring behavior separately across populations and conditions, and under deliberate shift or perturbation, rather than only in aggregate."@en ; skos:member , , . a skos:Collection ; skos:prefLabel "Evaluating generative & agentic systems"@en ; skos:notation "P3.4" ; skos:definition "Assessing open-ended outputs and multi-step runs where no single reference answer exists, using graders, sampling-aware metrics, and contamination controls."@en ; skos:member , , , , , , , , , . a skos:Collection ; skos:prefLabel "ML testing & debugging"@en ; skos:notation "P3.5" ; skos:definition "Testing data, models, pipelines, and their integration against stated expectations when exact expected outputs are often unknown."@en ; skos:member , , , , , , , . a skos:Collection ; skos:prefLabel "Deployment & serving"@en ; skos:notation "P4" ; skos:scopeNote "Serving architectures, latency and goodput, request scheduling, LLM inference mechanisms, structured generation, test-time compute, retrieval, and on-device deployment."@en ; skos:member , , , , , , , , . a skos:Collection ; skos:prefLabel "Serving architectures"@en ; skos:notation "P4.1" ; skos:definition "The structure through which predictions are produced and delivered, from offline batch scoring to online and serverless model servers, multi-model and multi-adapter hosting, and streaming speech services."@en ; skos:member , , , , , , , , , , , , . a skos:Collection ; skos:prefLabel "Compound AI systems & model routing"@en ; skos:notation "P4.2" ; skos:definition "Composing several models, retrievers, and tools into one application, and routing, cascading, caching, and gating requests across models to meet quality and cost targets."@en ; skos:member , , , , , . a skos:Collection ; skos:prefLabel "Latency, throughput & goodput"@en ; skos:notation "P4.3" ; skos:definition "Defining and meeting per-request latency and useful-throughput targets, including tail behavior and token-level latency for generative services."@en ; skos:member , , , , , , . a skos:Collection ; skos:prefLabel "Batching, admission & autoscaling"@en ; skos:notation "P4.4" ; skos:definition "Policies that group, admit, distribute, and provision for requests so a service meets its targets under varying load."@en ; skos:member , , , , , , , , . a skos:Collection ; skos:prefLabel "LLM & generative inference"@en ; skos:notation "P4.5" ; skos:definition "Mechanisms specific to generative inference: managing attention state, separating prompt processing from token generation, accelerating autoregressive decoding, and reducing iterative denoising cost."@en ; skos:member , , , , , , , , , , , , , , , . a skos:Collection ; skos:prefLabel "Decoding, sampling & structured output"@en ; skos:notation "P4.6" ; skos:definition "Controlling how tokens are selected and constraining generated output to required syntax or schemas."@en ; skos:member , , , . a skos:Collection ; skos:prefLabel "Test-time compute"@en ; skos:notation "P4.7" ; skos:definition "Spending additional inference computation on reasoning, sampling, search, and verification to raise output quality under a budget."@en ; skos:member , , , , , , . a skos:Collection ; skos:prefLabel "Retrieval & vector search"@en ; skos:notation "P4.8" ; skos:definition "Selecting relevant records from an external corpus by lexical, dense, or hybrid search and supplying them to a model at inference."@en ; skos:member , , , , , , , , , , . a skos:Collection ; skos:prefLabel "On-device & edge deployment"@en ; skos:notation "P4.9" ; skos:definition "Converting, packaging, executing, and updating models on phones, embedded devices, and microcontrollers under memory, energy, and connectivity limits."@en ; skos:member , , , , , . a skos:Collection ; skos:prefLabel "MLOps & operations"@en ; skos:notation "P5" ; skos:scopeNote "Release management, monitoring, drift and retraining, incident response, pipelines and reproducibility, platforms and teams, and retirement."@en ; skos:member , , , , , , . a skos:Collection ; skos:prefLabel "Versioning, rollout & rollback"@en ; skos:notation "P5.1" ; skos:definition "Identifying, gating, and progressively exposing model releases, with the ability to return to a known-good configuration."@en ; skos:member , , , , , , , . a skos:Collection ; skos:prefLabel "Monitoring & observability"@en ; skos:notation "P5.2" ; skos:definition "Instrumenting deployed systems and detecting, from logged signals, changes in inputs, outputs, quality, and health."@en ; skos:member , , , , , , . a skos:Collection ; skos:prefLabel "Drift, feedback loops & retraining"@en ; skos:notation "P5.3" ; skos:definition "Understanding how deployment data and outcomes change, including changes the system itself causes, and deciding when and how to update the model."@en ; skos:member , , , , , , , . a skos:Collection ; skos:prefLabel "Incident response"@en ; skos:notation "P5.4" ; skos:definition "Detecting, triaging, mitigating, and learning from production failures, including attributing them to data, model, or infrastructure causes."@en ; skos:member , , . a skos:Collection ; skos:prefLabel "ML pipelines & reproducibility"@en ; skos:notation "P5.5" ; skos:definition "Automating the steps that produce models and controlling their environments and configuration so runs can be repeated."@en ; skos:member , , , , , . a skos:Collection ; skos:prefLabel "ML platforms, teams & technical debt"@en ; skos:notation "P5.6" ; skos:definition "Shared internal platforms, team structures, and debt management that let an organization operate many models at sustainable cost."@en ; skos:member , , , , , , , , . a skos:Collection ; skos:prefLabel "Model retirement"@en ; skos:notation "P5.7" ; skos:definition "Ending a model's service by migrating its consumers, archiving what must be kept, and deleting data and resources that are no longer justified."@en ; skos:member , , , . a skos:Collection ; skos:prefLabel "Qualities"@en ; skos:notation "Q" ; skos:member , , , , , . a skos:Collection ; skos:prefLabel "Performance, efficiency & benchmarking"@en ; skos:notation "Q1" ; skos:scopeNote "Measuring and benchmarking where time, memory, and energy go, and the cross-layer techniques whose primary purpose is to reduce them, such as compression and hardware-aware design."@en ; skos:member , , , . a skos:Collection ; skos:prefLabel "Profiling & utilization"@en ; skos:notation "Q1.1" ; skos:definition "Measuring where execution time and resources go under a stated workload, and expressing how much of the hardware's capacity does useful model work."@en ; skos:member , , , , , . a skos:Collection ; skos:prefLabel "System benchmarking"@en ; skos:notation "Q1.2" ; skos:definition "Measuring training and inference system performance under specified workloads, quality targets, load scenarios, and run rules so results are comparable."@en ; skos:member , , , , , . a skos:Collection ; skos:prefLabel "Model compression"@en ; skos:notation "Q1.3" ; skos:definition "Techniques that shrink a trained or training model's parameters, precision, or rank to reduce memory, compute, and latency at a measured cost in quality."@en ; skos:member , , , , , , , , , , , , . a skos:Collection ; skos:prefLabel "Hardware-aware model design"@en ; skos:notation "Q1.4" ; skos:definition "Choosing or searching model architectures against a target device's measured cost, so that quality and efficiency are optimized together rather than compressed after the fact."@en ; skos:member , , , . a skos:Collection ; skos:prefLabel "Reliability & fault tolerance"@en ; skos:notation "Q2" ; skos:scopeNote "Infrastructure and service fault tolerance, from failure assumptions and recovery semantics to fleet failure rates, checkpoint economics, and resilience testing."@en ; skos:member , , , . a skos:Collection ; skos:prefLabel "Failure models, retries & degradation"@en ; skos:notation "Q2.1" ; skos:definition "Declared failure assumptions and the service-level mechanisms (retries, idempotency, replication, circuit breaking, degradation) that keep requests correct and available when components fail."@en ; skos:member , , , , , , , . a skos:Collection ; skos:prefLabel "Checkpointing & elastic recovery"@en ; skos:notation "Q2.2" ; skos:definition "Persisting recoverable state for long-running jobs and choosing checkpoint frequency and recovery strategy to minimize lost work and restart cost under a given failure rate."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Hardware faults & silent data corruption"@en ; skos:notation "Q2.3" ; skos:definition "How hardware fails in a fleet, from fail-stop crashes to gray and silent faults, how often it fails, and how such faults are detected and contained."@en ; skos:member , , , , , , , . a skos:Collection ; skos:prefLabel "Fault injection & resilience testing"@en ; skos:notation "Q2.4" ; skos:definition "Deliberately introducing software, infrastructure, or hardware faults under controlled scope to verify that detection, recovery, and degradation work as designed."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Security & privacy"@en ; skos:notation "Q3" ; skos:scopeNote "Adversaries and defenses across data, model, artifact, and execution boundaries, and the leakage of information about people or records through learned systems."@en ; skos:member , , , . a skos:Collection ; skos:prefLabel "Threat models & adversarial attacks"@en ; skos:notation "Q3.1" ; skos:definition "Specifying adversaries against learned systems and the attacks they mount on training data, model inputs, and instruction boundaries, with the defenses aimed at each."@en ; skos:member , , , , , , , , , , . a skos:Collection ; skos:prefLabel "Model & supply-chain security"@en ; skos:notation "Q3.2" ; skos:definition "Protecting models as valuable artifacts from theft and tampering, and protecting the chain of datasets, pretrained weights, libraries, and builds that produce them."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Privacy leakage & differential privacy"@en ; skos:notation "Q3.3" ; skos:definition "Privacy requirements for data about people, the attacks that extract it from models and outputs, and formal or architectural mechanisms that bound or remove that leakage."@en ; skos:member , , , , , , , , , . a skos:Collection ; skos:prefLabel "Access control & isolation"@en ; skos:notation "Q3.4" ; skos:definition "Constraining who may act, what code may reach, and which parties can observe data in use, through authorization, least privilege, isolation, and hardware-rooted trust."@en ; skos:member , , , , , , , , , , . a skos:Collection ; skos:prefLabel "Safety & assurance"@en ; skos:notation "Q4" ; skos:scopeNote "Identifying hazards, bounding the operating domain, enforcing safety at runtime, verifying and validating behavior, and arguing acceptable residual risk against recognized standards."@en ; skos:member , , , . a skos:Collection ; skos:prefLabel "Hazard analysis & operating domains"@en ; skos:notation "Q4.1" ; skos:definition "Identifying how a system can contribute to harm, including hazards from functional insufficiency rather than faults, and fixing the operating conditions within which safety claims are made."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Runtime assurance & safety filters"@en ; skos:notation "Q4.2" ; skos:definition "Mechanisms that monitor a learned component at run time and override, constrain, or replace its outputs so the system stays within a verified safe set."@en ; skos:member , , , , , , . a skos:Collection ; skos:prefLabel "Safety verification & validation"@en ; skos:notation "Q4.3" ; skos:definition "Producing evidence that a system meets its safety requirements across its operating domain, through scenario testing, falsification, formal proof, and statistical validation of rare failures."@en ; skos:member , , , , , , . a skos:Collection ; skos:prefLabel "Safety cases, standards & certification"@en ; skos:notation "Q4.4" ; skos:definition "Structured arguments that a system is acceptably safe for a defined use, the standards that set what evidence is required, and the field indicators that keep the argument valid after release."@en ; skos:member , , , , , , . a skos:Collection ; skos:prefLabel "Responsible AI"@en ; skos:notation "Q5" ; skos:scopeNote "Measuring and mitigating unfair outcomes, explaining and documenting models and data, keeping people able to oversee and contest decisions, and governing AI risk under organizational and legal regimes."@en ; skos:member , , , . a skos:Collection ; skos:prefLabel "Fairness & bias"@en ; skos:notation "Q5.1" ; skos:definition "Specifying, measuring, and mitigating how a system's outcomes and errors differ across people and groups."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Explainability & documentation"@en ; skos:notation "Q5.2" ; skos:definition "Making model behavior understandable to the people who build, deploy, and are affected by it, through explanations, interpretable models, and standardized documentation of models and data."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Human oversight & contestability"@en ; skos:notation "Q5.3" ; skos:definition "Arrangements that let people supervise, intervene in, take over from, and challenge consequential system decisions, and the human-factors limits on doing so."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "AI governance, audit & regulation"@en ; skos:notation "Q5.4" ; skos:definition "Organizational and legal mechanisms that identify, assess, and control AI risk across a system's life, including risk-management frameworks, impact assessments, audits, and regulation."@en ; skos:member , , , , , , . a skos:Collection ; skos:prefLabel "Sustainability & cost"@en ; skos:notation "Q6" ; skos:scopeNote "Accounting for environmental and monetary consequences of ML systems under declared boundaries, and the decisions that reduce them per useful outcome."@en ; skos:member , , . a skos:Collection ; skos:prefLabel "Energy & carbon accounting"@en ; skos:notation "Q6.1" ; skos:definition "Measuring the energy a workload consumes and converting it to operational emissions under declared boundaries, and scheduling work to reduce both."@en ; skos:member , , , , , , , , , , , . a skos:Collection ; skos:prefLabel "Embodied carbon & lifecycle impact"@en ; skos:notation "Q6.2" ; skos:definition "Environmental impacts beyond operating energy, from manufacturing emissions and water to hardware lifetime and disposal, assessed over the full hardware lifecycle."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Cost, TCO & unit economics"@en ; skos:notation "Q6.3" ; skos:definition "Monetary accounting for ML systems, from total cost of ownership to cost per token or accepted outcome, and the architectural choices that lower cost at fixed quality."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "System types"@en ; skos:notation "T" ; skos:member , . a skos:Collection ; skos:prefLabel "Agentic systems"@en ; skos:notation "T1" ; skos:scopeNote "Context engineering and agent memory, tool and action interfaces, the agent harness and its control loop, durable agent state, and multi-agent coordination."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Context engineering & agent memory"@en ; skos:notation "T1.1" ; skos:definition "Deciding what information enters each model call over a long task, and maintaining persistent memory the agent can write, consolidate, and retrieve across calls and sessions."@en ; skos:member , , , . a skos:Collection ; skos:prefLabel "Tool use & action interfaces"@en ; skos:notation "T1.2" ; skos:definition "The interfaces through which a model acts on the world, including structured tool calls, tool discovery protocols, computer use, and the effects and authority those actions carry."@en ; skos:member , , , , , . a skos:Collection ; skos:prefLabel "Agent harness & control loop"@en ; skos:notation "T1.3" ; skos:definition "The runtime that owns an agent's loop of model calls and tool actions, enforcing budgets, stopping rules, and approval gates and recording the trajectory."@en ; skos:member , , , , , , , . a skos:Collection ; skos:prefLabel "Durable execution & agent state"@en ; skos:notation "T1.4" ; skos:definition "Persisting an agent's progress so long-running trajectories survive crashes, pauses, and redeployments without repeating completed effects, and can be replayed for debugging."@en ; skos:member , , . a skos:Collection ; skos:prefLabel "Multi-agent orchestration"@en ; skos:notation "T1.5" ; skos:definition "Decomposing a task across several model-driven agents and coordinating their delegation, shared state, and results, including when coordination costs outweigh the benefit."@en ; skos:member , , , . a skos:Collection ; skos:prefLabel "Embodied & physical systems"@en ; skos:notation "T2" ; skos:scopeNote "Sensing and perception, state estimation and mapping, motion planning and control, real-time and mixed-criticality computing, and robot learning from demonstration and simulation."@en ; skos:member , , , , . a skos:Collection ; skos:prefLabel "Sensing, perception & fusion"@en ; skos:notation "T2.1" ; skos:definition "Turning time-stamped measurements from physical sensors into task-relevant estimates of the scene, including calibration, synchronization, and fusion across modalities."@en ; skos:member , , , , , . a skos:Collection ; skos:prefLabel "State estimation, localization & mapping"@en ; skos:notation "T2.2" ; skos:definition "Inferring a robot's and its environment's latent state and uncertainty over time, through recursive filtering, localization and mapping, and learned predictive world models."@en ; skos:member , , , , , . a skos:Collection ; skos:prefLabel "Motion planning & control"@en ; skos:notation "T2.3" ; skos:definition "Generating feasible, collision-free motions and closing feedback loops that drive a physical system along them under dynamics, contact, and actuation limits."@en ; skos:member , , , , , , , . a skos:Collection ; skos:prefLabel "Real-time & mixed-criticality computing"@en ; skos:notation "T2.4" ; skos:definition "Guaranteeing that sensing, inference, and control complete within deadlines on shared embedded hardware, and isolating safety-critical functions from less trusted ones."@en ; skos:member , , , , , , , , , . a skos:Collection ; skos:prefLabel "Robot learning & sim-to-real"@en ; skos:notation "T2.5" ; skos:definition "Learning robot policies from teleoperated demonstrations, simulation, and pretrained vision-language backbones, and transferring them to physical hardware."@en ; skos:member , , , , , , . a skos:Concept ; skos:inScheme ; skos:prefLabel "System requirements"@en ; skos:definition "Measurable desired behavior and constraints for a specified workload."@en ; mlsys:determines ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Baseline"@en ; skos:definition "A reference method or system used to justify the value of a proposed change."@en ; mlsys:part-of ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Proxy objective"@en ; skos:definition "A measurable target used in place of the outcome actually desired."@en ; mlsys:measures ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Service-level objective"@en ; skos:altLabel "SLO"@en ; skos:definition "A target for a defined service measure over a specified population and window."@en ; mlsys:determines ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Workload characterization"@en ; skos:definition "The input, request, model, and execution distributions that a system must support."@en ; mlsys:determines ; mlsys:uses ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Execution placement"@en ; skos:definition "The choice of where system components execute, such as device, edge, or datacenter."@en ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data quality"@en ; skos:definition "Fitness of recorded examples for the intended task, including validity, coverage, and error."@en ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Label quality"@en ; skos:definition "Correctness, consistency, and availability of supervisory targets."@en ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Feature engineering"@en ; skos:definition "Constructing model inputs whose semantics and availability match the task."@en ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Missing data"@en ; skos:definition "Absent observations whose pattern and treatment can change learned behavior."@en ; mlsys:determines ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data leakage"@en ; skos:definition "Information unavailable at the intended prediction boundary entering development or evaluation."@en ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Point-in-time correctness"@en ; skos:definition "Constructing examples using only information available at their stated prediction time."@en ; mlsys:mitigates ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data lineage"@en ; skos:definition "Recorded dependencies that identify the sources and transformations of a dataset or artifact."@en ; mlsys:detects ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Training-serving skew"@en ; skos:definition "A mismatch between the data transformations or feature semantics used in training and serving."@en ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Dataset deduplication"@en ; skos:definition "Identifying and managing repeated or near-duplicate examples."@en ; mlsys:part-of ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data selection"@en ; skos:definition "Allocating learning effort to examples according to value, coverage, and acquisition cost."@en ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Tokenization"@en ; skos:definition "Mapping text into vocabulary-dependent symbols and their identifiers."@en ; mlsys:determines ; mlsys:determines ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Distribution shift"@en ; skos:definition "A difference between development and deployment data-generating distributions."@en ; mlsys:reduces ; mlsys:causes ; mlsys:causes ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Linear models"@en ; skos:definition "Models whose score is linear in their supplied features, with a possible output link function."@en ; mlsys:uses ; mlsys:determines ; mlsys:alternative-to ; mlsys:specializes ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Tree ensembles"@en ; skos:definition "Collections of decision trees whose branch traversal and aggregation determine predictions."@en ; mlsys:determines ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Kernel methods"@en ; skos:definition "Models using similarity functions whose storage and evaluation costs depend on their representation."@en ; mlsys:determines ; mlsys:increases ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Nearest-neighbor models"@en ; skos:definition "Predictions or retrieval based on similarity to stored examples or vectors."@en ; mlsys:uses ; mlsys:determines ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Neural networks"@en ; skos:definition "Parameterized compositions of differentiable operations used to learn representations and predictions."@en ; mlsys:uses ; mlsys:determines ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Inductive bias"@en ; skos:definition "Structural assumptions that favor some learned functions over others."@en ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Embeddings"@en ; skos:definition "Vector representations whose dimension and lookup or construction method affect storage and computation."@en ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Attention"@en ; skos:definition "Input-dependent aggregation of value vectors using query-key compatibility scores."@en ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Position encoding"@en ; skos:altLabel "Positional encoding"@en ; skos:definition "Representations that make sequence position available to a model."@en ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Training loss"@en ; skos:definition "A computable criterion optimized on training examples."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Automatic differentiation"@en ; skos:definition "Computing derivatives by composing local derivative rules through an execution graph."@en ; mlsys:enables ; mlsys:uses ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Optimization algorithm"@en ; skos:definition "A rule that uses training information to update model parameters and optimizer state."@en ; mlsys:requires ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Regularization"@en ; skos:definition "Constraints or training procedures used to control sensitivity and generalization behavior."@en ; mlsys:determines ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Training batch size"@en ; skos:definition "The number of examples contributing to an optimization update or microstep."@en ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Training state"@en ; skos:definition "The parameters, gradients, optimizer variables, and other state needed to continue training."@en ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Activation recomputation"@en ; skos:altLabel "Activation checkpointing"@en ; skos:altLabel "Gradient checkpointing"@en ; skos:definition "Recreating forward intermediates during backward execution to reduce retained memory."@en ; mlsys:reduces ; mlsys:increases ; mlsys:alternative-to ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Empirical scaling laws"@en ; skos:definition "Fitted relationships between resources, data, model size, and observed model performance."@en ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Hyperparameter search"@en ; skos:altLabel "Hyperparameter optimization"@en ; skos:altLabel "HPO"@en ; skos:definition "Choosing model or training settings through a sequence of evaluations."@en ; mlsys:causes ; mlsys:uses ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Experiment-budget allocation"@en ; skos:definition "Dividing a finite evaluation budget across configurations, repetitions, and fidelities."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Transfer learning"@en ; skos:definition "Adapting representations or parameters learned from one setting to another."@en ; mlsys:reduces ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Parameter-efficient adaptation"@en ; skos:altLabel "PEFT"@en ; skos:definition "Adapting a model by changing a limited parameter subset or added parameterization."@en ; mlsys:implements ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Tensor layout"@en ; skos:definition "The mapping from tensor indices to storage locations, including strides and materialization."@en ; mlsys:determines ; mlsys:part-of ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Matrix multiplication"@en ; skos:definition "A dense linear-algebra operation whose tiling, shape, and precision affect execution cost."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Numerical precision"@en ; skos:definition "The representable values and arithmetic behavior of a numerical format."@en ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Numerical stability"@en ; skos:definition "Sensitivity of a computed result to finite-precision operations and intermediate representations."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Mixed-precision training"@en ; skos:definition "Training with multiple numerical formats and explicit handling of their range and accumulation behavior."@en ; mlsys:requires ; mlsys:reduces ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Quantization"@en ; skos:definition "Representing values using a reduced discrete set with a specified encoding and approximation rule."@en ; mlsys:reduces ; mlsys:requires ; mlsys:reduces ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Sparsity"@en ; skos:definition "A pattern of absent or zero values whose execution benefit depends on representation and support."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Knowledge distillation"@en ; skos:definition "Training a student model using signals from a teacher model or ensemble."@en ; mlsys:produces ; mlsys:reduces ; mlsys:enables ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Graph compilation"@en ; skos:definition "Transforming a model or execution graph into an executable representation."@en ; mlsys:enables ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Kernel fusion"@en ; skos:definition "Combining compatible operations to reduce intermediate materialization or dispatch."@en ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Host dispatch overhead"@en ; skos:altLabel "kernel launch overhead"@en ; skos:definition "Host-side work required to launch or orchestrate device operations."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Memory capacity"@en ; skos:definition "Available storage for concurrently resident state and intermediates."@en ; mlsys:limits ; mlsys:limits ; mlsys:limits ; mlsys:limits ; mlsys:limits ; mlsys:limits ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Memory bandwidth"@en ; skos:definition "The sustainable rate at which bytes move across a specified memory interface."@en ; mlsys:determines ; mlsys:limits ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Compute throughput"@en ; skos:definition "The sustainable rate of a specified arithmetic workload on an execution resource."@en ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Arithmetic intensity"@en ; skos:definition "Arithmetic work per byte transferred at a specified memory boundary."@en ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Roofline bound"@en ; skos:definition "A performance bound combining a compute ceiling with an intensity-dependent bandwidth ceiling."@en ; mlsys:limits ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data locality"@en ; skos:definition "Reuse of data near its consumers across time or execution locations."@en ; mlsys:increases ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Tiling"@en ; skos:definition "Partitioning computation so working data can be reused within a chosen storage level."@en ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Profiling"@en ; skos:definition "Measuring execution to attribute time, work, stalls, and resource use."@en ; mlsys:measures ; mlsys:measures ; mlsys:measures ; mlsys:enables ; mlsys:produces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Resource utilization"@en ; skos:definition "The fraction of a resource's defined service capacity that is occupied over a stated interval."@en ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Serial fraction"@en ; skos:definition "The portion of execution that a proposed parallel or accelerated transformation leaves unchanged."@en ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Collective communication"@en ; skos:definition "Coordinated data exchange and reduction across a process group."@en ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data parallelism"@en ; skos:definition "Executing model replicas on different examples while coordinating learning state."@en ; mlsys:uses ; mlsys:increases ; mlsys:uses ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Tensor parallelism"@en ; skos:definition "Partitioning tensor operations and their state across devices."@en ; mlsys:requires ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Pipeline parallelism"@en ; skos:definition "Partitioning a computation into stages that process different microbatches concurrently."@en ; mlsys:produces ; mlsys:reduces ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Training-state sharding"@en ; skos:altLabel "ZeRO"@en ; skos:altLabel "FSDP"@en ; skos:altLabel "fully sharded data parallel"@en ; skos:definition "Partitioning parameters, gradients, or optimizer state across workers."@en ; mlsys:reduces ; mlsys:uses ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Communication overlap"@en ; skos:definition "Scheduling communication concurrently with independent computation."@en ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Network topology"@en ; skos:definition "The connectivity and capacity structure that constrains communication paths."@en ; mlsys:limits ; mlsys:determines ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Stragglers"@en ; skos:definition "Workers or tasks whose delayed completion holds back coordinated work."@en ; mlsys:limits ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Recovery checkpoint"@en ; skos:definition "A persisted state from which a specified computation can resume."@en ; mlsys:requires ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Consistent checkpoint cut"@en ; skos:definition "A recoverable collection of distributed states with compatible dependencies."@en ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Replicated-log consensus"@en ; skos:altLabel "Paxos"@en ; skos:altLabel "Raft"@en ; skos:definition "Agreement on an ordered state history under specified timing and failure assumptions."@en ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Fencing"@en ; skos:definition "Rejecting effects from an actor whose epoch or authority is stale."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Federated learning"@en ; skos:definition "Coordinating model learning across separately controlled data holders."@en ; mlsys:requires ; mlsys:uses ; mlsys:requires ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Arrival rate"@en ; skos:definition "The frequency and distribution of requests reaching a service boundary."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Service capacity"@en ; skos:definition "The completion rate supported by a service under a stated workload and operating point."@en ; mlsys:limits ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Queueing delay"@en ; skos:definition "Time a request waits for a resource before receiving service."@en ; mlsys:increases ; mlsys:determines ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Tail latency"@en ; skos:definition "A high quantile of a defined request-latency distribution."@en ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Request batching"@en ; skos:definition "Executing requests together to share work or improve resource use."@en ; mlsys:increases ; mlsys:increases ; mlsys:reduces ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Admission control"@en ; skos:definition "Accepting, delaying, or rejecting work according to capacity, policy, and service constraints."@en ; mlsys:requires ; mlsys:limits ; mlsys:implements . a skos:Concept ; skos:inScheme ; skos:prefLabel "Autoscaling"@en ; skos:definition "Changing provisioned service resources in response to measured or predicted demand."@en ; mlsys:determines ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "KV cache"@en ; skos:altLabel "Key-value cache"@en ; skos:definition "Retained attention keys and values reused during autoregressive execution."@en ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Prefill and decode"@en ; skos:definition "The prompt-processing and incremental token-generation phases of autoregressive inference."@en ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Paged KV allocation"@en ; skos:altLabel "PagedAttention"@en ; skos:definition "Managing attention state in blocks to reduce allocation waste and support sharing."@en ; mlsys:implements ; mlsys:enables ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Prefix caching"@en ; skos:definition "Reusing compatible computed state for an identical request prefix."@en ; mlsys:uses ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Speculative decoding"@en ; skos:definition "Using proposed tokens and a validating procedure to accelerate generation while preserving a specified output distribution."@en ; mlsys:increases ; mlsys:reduces ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "SLO goodput"@en ; skos:definition "The rate of useful completed work that satisfies the specified quality and service criteria."@en ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Generalization"@en ; skos:definition "Performance on relevant examples or conditions beyond those used to fit the system."@en ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Evaluation split"@en ; skos:definition "A partitioning protocol that separates learning, selection, and final assessment."@en ; mlsys:measures ; mlsys:mitigates ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model-selection optimism"@en ; skos:definition "Overstated performance caused by choosing a winner using noisy evaluation results."@en ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Evaluation metrics"@en ; skos:definition "Specified measurements of predictive or task behavior, with explicit populations and aggregation."@en ; mlsys:measures ; mlsys:determines ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Probability calibration"@en ; skos:definition "Agreement between predicted probabilities and observed event frequencies under an evaluation distribution."@en ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Statistical uncertainty"@en ; skos:definition "Uncertainty in an estimated quantity due to finite and dependent observations."@en ; mlsys:limits ; mlsys:limits ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Benchmark protocol"@en ; skos:definition "A reproducible workload, measurement boundary, and comparison procedure."@en ; mlsys:measures ; mlsys:requires ; mlsys:requires ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Online experimentation"@en ; skos:definition "A controlled intervention used to estimate effects on deployed-system outcomes."@en ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Sliced evaluation"@en ; skos:definition "Measuring behavior separately on relevant populations, conditions, or workload groups."@en ; mlsys:measures ; mlsys:detects . a skos:Concept ; skos:inScheme ; skos:prefLabel "Selective prediction"@en ; skos:definition "Predicting on admitted cases and abstaining or deferring the rest."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Artifact versioning"@en ; skos:definition "Identifying the exact model, code, data, configuration, and dependencies in an execution or release."@en ; mlsys:enables ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Production monitoring"@en ; skos:definition "Collecting and interpreting signals about deployed system behavior and health."@en ; mlsys:detects ; mlsys:measures ; mlsys:produces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Staged rollout"@en ; skos:definition "Increasing exposure to a candidate system through bounded release stages."@en ; mlsys:requires ; mlsys:uses ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Rollback"@en ; skos:definition "Returning to a previously accepted deployment configuration with compatible state."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Deployment feedback loop"@en ; skos:definition "A deployed system's outputs changing the data or behavior encountered by future learning or evaluation."@en ; mlsys:reduces ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Retraining"@en ; skos:definition "Updating learned parameters from new or revised training information."@en ; mlsys:mitigates ; mlsys:requires ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Incident response"@en ; skos:definition "Detecting, containing, diagnosing, and recovering from unexpected system behavior."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "System observability"@en ; skos:definition "Instrumentation that permits investigation of internal execution from recorded outputs, traces, and metrics."@en ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Threat modeling"@en ; skos:definition "Specifying adversaries, assets, interfaces, capabilities, and failure consequences."@en ; mlsys:determines ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data poisoning"@en ; skos:definition "Adversarial manipulation of learning data or signals to alter learned behavior."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Adversarial evasion"@en ; skos:definition "Adversarially chosen inference inputs intended to cause harmful model behavior."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Privacy requirements"@en ; skos:definition "Requirements governing the collection, use, retention, and disclosure of information about people or records."@en ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Differential privacy"@en ; skos:altLabel "DP"@en ; skos:definition "A parameterized bound on changes in output distributions under a specified neighboring-dataset relation."@en ; mlsys:implements ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Fairness criteria"@en ; skos:definition "Explicit requirements for how system outcomes or errors are distributed across affected groups."@en ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Authorization"@en ; skos:definition "Determining whether an actor may perform a specified action on a resource."@en ; mlsys:mitigates ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Execution isolation"@en ; skos:definition "Restricting which resources and effects a computation can access."@en ; mlsys:implements ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Human oversight"@en ; skos:definition "An arrangement in which people can understand, intervene in, and contest consequential system behavior."@en ; mlsys:requires ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Operating energy"@en ; skos:definition "Energy consumed by a specified system boundary over a defined workload or interval."@en ; mlsys:increases ; mlsys:increases ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Lifecycle carbon"@en ; skos:definition "Emissions attributed to operation and embodied system resources under declared accounting boundaries."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Cost per useful outcome"@en ; skos:definition "Total relevant expenditure divided by outcomes meeting a specified acceptance criterion."@en ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Capacity planning"@en ; skos:definition "Provisioning resources for a workload distribution and service objectives."@en ; mlsys:requires ; mlsys:uses ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Agent trajectory"@en ; skos:definition "A sequence of observations, model proposals, tool actions, and external outcomes for a task."@en ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Agent planning"@en ; skos:definition "Constructing and revising prospective actions in response to goals, dependencies, and observations."@en ; mlsys:determines ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Context window"@en ; skos:altLabel "context length"@en ; skos:altLabel "sequence length"@en ; skos:definition "The information serialized into a model invocation under its input budget."@en ; mlsys:limits ; mlsys:limits ; mlsys:increases ; mlsys:determines ; mlsys:limits ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Retrieval"@en ; skos:definition "Selecting relevant records from an external collection using an explicit query and ranking process."@en ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Approximate nearest-neighbor search"@en ; skos:altLabel "ANN"@en ; skos:definition "Trading exact neighbor recovery for reduced search cost using an index or approximation."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Hallucination"@en ; skos:altLabel "unsupported generated claims"@en ; skos:definition "Generated factual assertions or summaries not supported by the relevant source or evidence."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Tool contract"@en ; skos:definition "A specification of tool inputs, outputs, authority, effects, and failure behavior."@en ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Idempotency"@en ; skos:definition "A specified operation's repeated execution having the same intended effect as one execution."@en ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Compensating action"@en ; skos:definition "An explicit action intended to remedy an earlier effect whose direct reversal is unavailable."@en ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Prompt injection"@en ; skos:definition "Untrusted content attempting to redirect model-driven behavior across an instruction or authority boundary."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Trajectory evaluation"@en ; skos:definition "Assessing task outcomes, intermediate effects, and constraint compliance across a complete agent run."@en ; mlsys:measures ; mlsys:uses ; mlsys:measures ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Reward hacking"@en ; skos:definition "Improving a measured reward through behavior that fails the intended task objective."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Physical dynamics"@en ; skos:definition "The evolution of a physical system under inputs, disturbances, and constraints."@en ; mlsys:limits ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Sensor calibration"@en ; skos:definition "Establishing how observations relate to physical quantities, coordinates, and time."@en ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "State estimation"@en ; skos:definition "Inferring latent physical state and uncertainty from observations and a model."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "State observability"@en ; skos:definition "Whether distinct internal states can be distinguished through the available observations and model."@en ; mlsys:limits ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Observation age"@en ; skos:definition "Elapsed time between a physical observation and its use in a decision."@en ; mlsys:determines ; mlsys:requires ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Feedback delay"@en ; skos:definition "Delay around a sensing, computation, communication, and actuation loop."@en ; mlsys:limits ; mlsys:increases ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Worst-case execution time"@en ; skos:altLabel "WCET"@en ; skos:definition "A justified execution-time upper bound under specified hardware and execution assumptions."@en ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Feedback control"@en ; skos:definition "Selecting inputs using measured or estimated deviation from desired behavior."@en ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Trajectory feasibility"@en ; skos:definition "Satisfaction of motion, state, input, and dynamic constraints along a proposed trajectory."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Contact and compliance"@en ; skos:definition "Mechanical interaction whose forces depend on geometry, motion, stiffness, and damping."@en ; mlsys:determines ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Simulation-to-real transfer"@en ; skos:definition "Deployment of a learned policy under differences between simulated and physical conditions."@en ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Runtime assurance"@en ; skos:definition "Monitoring and constraining execution using explicitly stated safety premises and fallback mechanisms."@en ; mlsys:limits ; mlsys:uses ; mlsys:mitigates ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Stopping envelope"@en ; skos:altLabel "safe stopping region"@en ; skos:definition "The state and clearance conditions within which a specified stopping response remains feasible."@en ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Operational design domain"@en ; skos:altLabel "ODD"@en ; skos:definition "The conditions under which a physical system's behavior and evidence are intended to apply."@en ; mlsys:limits ; mlsys:determines ; mlsys:limits ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Safety case"@en ; skos:definition "A structured argument connecting claims, evidence, assumptions, and defeaters for a defined system."@en ; mlsys:determines ; mlsys:requires ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Structured output constraint"@en ; skos:definition "Restricting the syntax or schema of model outputs without establishing factual correctness or authority."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data acquisition and sampling"@en ; skos:definition "The observation and selection process that determines which examples and populations enter a dataset."@en ; mlsys:determines ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data contract"@en ; skos:definition "A specification of schema, semantics, validity rules, and availability for data exchanged between components."@en ; mlsys:determines ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Delayed outcomes"@en ; skos:definition "Target outcomes that become observable only after a delay or for a selected subset of cases."@en ; mlsys:limits ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Decision policy"@en ; skos:definition "The rule that converts model outputs and system state into actions, thresholds, abstentions, or escalation."@en ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Robustness evaluation"@en ; skos:definition "Testing behavior under explicitly chosen shifts, perturbations, faults, or operating stresses."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Supervised learning"@en ; skos:definition "Learning a prediction rule from examples with target outputs."@en ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Self-supervised learning"@en ; skos:definition "Learning from targets constructed from the observed data itself."@en ; mlsys:determines ; mlsys:enables ; mlsys:produces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Unsupervised learning"@en ; skos:definition "Learning structure in data without externally supplied example-level targets."@en ; mlsys:implements . a skos:Concept ; skos:inScheme ; skos:prefLabel "Reinforcement learning"@en ; skos:definition "Learning a policy from action-dependent observations and reward signals."@en ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Imitation learning"@en ; skos:definition "Learning a policy from demonstrated behavior and possibly corrective interaction."@en ; mlsys:causes ; mlsys:uses ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model artifact"@en ; skos:definition "The learned parameters, structures, example stores, and metadata required to execute a model."@en ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Failure model"@en ; skos:definition "The declared classes of faults and timing assumptions under which a mechanism is analyzed."@en ; mlsys:limits ; mlsys:determines ; mlsys:limits ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Backpressure"@en ; skos:definition "A mechanism that signals downstream saturation to limit or slow upstream production."@en ; mlsys:limits ; mlsys:limits ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Privacy leakage"@en ; skos:definition "Information about people or records revealed through data, artifacts, or model interaction."@en ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Execution time"@en ; skos:definition "Elapsed time for a specified execution boundary, workload, and operating condition."@en ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Irreversible effects"@en ; skos:definition "External consequences that cannot be fully undone by restoring software state, retrying, or performing a compensating action."@en ; mlsys:limits ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Dataset storage layout"@en ; skos:definition "Organization of stored dataset records, columns, partitions, and metadata that determines which data an access reads or decodes."@en ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data ingestion"@en ; skos:definition "Reading externally produced records into a processing or learning system under declared delivery, ordering, and timing assumptions."@en ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Schema evolution"@en ; skos:definition "Changing a data schema while explicitly managing how existing data and producers or consumers remain interpretable and compatible."@en ; mlsys:requires ; mlsys:causes ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Test oracle"@en ; skos:definition "A specification, procedure, reference, or informed judgment used to assess whether observed behavior satisfies stated expectations for a test."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Integration testing"@en ; skos:definition "Testing interactions among assembled components against stated interface and behavior expectations."@en ; mlsys:requires ; mlsys:detects ; mlsys:measures ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Failure injection"@en ; skos:definition "Deliberately introducing selected faults or disruptive conditions to observe system response under controlled experimental boundaries."@en ; mlsys:measures ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Dependency environment"@en ; skos:definition "The software packages, versions, native libraries, and runtime configuration on which an execution depends."@en ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Workflow orchestration"@en ; skos:definition "Coordinating dependent computation steps, their data or control flow, execution conditions, and failure handling."@en ; mlsys:produces ; mlsys:requires ; mlsys:implements . a skos:Concept ; skos:inScheme ; skos:prefLabel "Experiment reproducibility"@en ; skos:definition "The ability to repeat an experimental procedure and obtain results consistent with declared criteria when its inputs, environment, randomness, and execution conditions are specified."@en ; mlsys:requires ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Accelerator execution architecture"@en ; skos:altLabel "GPU architecture"@en ; skos:altLabel "streaming multiprocessor"@en ; skos:altLabel "SM"@en ; skos:definition "Organization of arithmetic units, execution lanes or threads, scheduling resources, and local storage that determines how supported operations run. SIMD/SIMT, matrix units, and systolic arrays are distinct examples."@en ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Memory hierarchy"@en ; skos:definition "Arrangement of storage levels and transfer paths with differing capacity, latency, bandwidth, access control, and management mechanisms. The available levels depend on the platform."@en ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Host-device interconnect"@en ; skos:definition "Communication path through which a host and accelerator exchange data and control, with transport, topology, and memory-access semantics that affect transfers."@en ; mlsys:increases ; mlsys:limits ; mlsys:limits ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Power delivery and integrity"@en ; skos:definition "Supply paths, conversion stages, limits, and transient behavior that must maintain required voltages while workloads and actuators draw or return power."@en ; mlsys:limits ; mlsys:limits ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Thermal behavior and throttling"@en ; skos:definition "Evolution of temperature under heat generation and removal, and the performance controls or operating restrictions that follow under a platform thermal policy."@en ; mlsys:increases ; mlsys:increases ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Service-level indicator"@en ; skos:altLabel "SLI"@en ; skos:definition "A quantitative measure of one aspect of delivered service, such as latency, availability, or accepted-output rate, computed over a stated population and window."@en ; mlsys:part-of ; mlsys:measures ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Error budget"@en ; skos:definition "The amount of service-level-objective violation permitted over a window, used to pace changes and risk against a reliability target."@en ; mlsys:determines ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Latency budget"@en ; skos:definition "An allocation of an end-to-end latency target across the stages of a request path, such as network, preprocessing, inference, and postprocessing."@en ; mlsys:limits ; mlsys:limits ; mlsys:limits ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Operation and byte accounting"@en ; skos:altLabel "FLOP counting"@en ; skos:altLabel "resource accounting"@en ; skos:definition "Estimating a workload's arithmetic operations and bytes moved per example, token, or step from model shapes, to predict cost before measurement."@en ; mlsys:measures ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Memory footprint"@en ; skos:definition "The peak memory a workload occupies, summed over parameters, activations, optimizer or cache state, and runtime buffers."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Workload trace"@en ; skos:definition "A recorded sequence of requests or jobs with arrival times, sizes, and durations, replayed to evaluate capacity, scheduling, or serving policies."@en ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Cost-sensitive decision"@en ; skos:definition "Choosing actions or thresholds to minimize expected cost when different error types carry unequal consequences."@en ; mlsys:determines ; mlsys:requires ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Pareto frontier"@en ; skos:definition "The set of configurations for which no objective, such as accuracy, latency, or cost, can improve without worsening another."@en ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Transformer"@en ; skos:definition "A neural architecture of stacked attention and feed-forward blocks with residual connections and normalization, processing all sequence positions in parallel during training."@en ; mlsys:uses ; mlsys:uses ; mlsys:uses ; mlsys:specializes ; mlsys:produces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Deep recommendation model"@en ; skos:altLabel "DLRM"@en ; skos:definition "Models combining large sparse embedding tables for categorical features with dense layers, whose cost is often dominated by embedding lookups and memory capacity."@en ; mlsys:uses ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Diffusion model"@en ; skos:definition "A generative model that learns to reverse a gradual noising process, producing samples through many iterative denoising network evaluations."@en ; mlsys:increases ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Convolution"@en ; skos:altLabel "CNN"@en ; skos:definition "A locally connected operation applying shared filters across spatial or temporal positions, giving high weight reuse and arithmetic intensity."@en ; mlsys:increases ; mlsys:implements . a skos:Concept ; skos:inScheme ; skos:prefLabel "Recurrent computation"@en ; skos:altLabel "RNN"@en ; skos:altLabel "LSTM"@en ; skos:definition "Computation that updates a hidden state step by step over a sequence, creating sequential dependencies that limit parallelism across time."@en ; mlsys:increases ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Normalization layers"@en ; skos:altLabel "BatchNorm"@en ; skos:altLabel "LayerNorm"@en ; skos:altLabel "RMSNorm"@en ; skos:definition "Layers that rescale activations using batch, layer, or root-mean-square statistics to stabilize and speed training of deep networks."@en ; mlsys:increases ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Mixture of experts"@en ; skos:altLabel "MoE"@en ; skos:definition "An architecture in which a learned router activates a small subset of expert subnetworks per token, decoupling total parameter count from per-token compute."@en ; mlsys:reduces ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Expert routing"@en ; skos:altLabel "gating"@en ; skos:altLabel "router"@en ; skos:definition "The gating function that assigns each token to one or more experts, such as top-k softmax gating or expert-choice routing."@en ; mlsys:determines ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Expert capacity"@en ; skos:altLabel "capacity factor"@en ; skos:definition "The maximum number of tokens an expert processes per batch, set by a capacity factor; tokens beyond it are dropped or rerouted."@en ; mlsys:limits ; mlsys:part-of ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Load-balancing loss"@en ; skos:altLabel "auxiliary loss"@en ; skos:definition "An auxiliary training objective that penalizes uneven token assignment across experts so routing stays balanced."@en ; mlsys:determines ; mlsys:mitigates ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Active parameters"@en ; skos:definition "The parameters actually used to process one token or example, as distinct from total stored parameters in conditionally computed models."@en ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Early exit"@en ; skos:altLabel "adaptive computation"@en ; skos:definition "Terminating inference at an intermediate layer once a confidence criterion is met, so per-input compute varies with input difficulty."@en ; mlsys:reduces ; mlsys:requires ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Sparse and sliding-window attention"@en ; skos:altLabel "local attention"@en ; skos:definition "Attention restricted to local windows or fixed sparse patterns so cost grows linearly rather than quadratically with sequence length."@en ; mlsys:enables ; mlsys:specializes ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Linear attention"@en ; skos:definition "Attention reformulated with kernel feature maps so it is computed in linear time with a fixed-size recurrent state."@en ; mlsys:reduces ; mlsys:specializes ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "State-space sequence model"@en ; skos:altLabel "SSM"@en ; skos:altLabel "Mamba"@en ; skos:altLabel "S4"@en ; skos:definition "Sequence models built on discretized linear state-space recurrences, trainable in parallel and run with constant-size state per step at inference."@en ; mlsys:specializes ; mlsys:alternative-to ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Grouped-query attention"@en ; skos:altLabel "GQA"@en ; skos:altLabel "multi-query attention"@en ; skos:altLabel "MQA"@en ; skos:definition "Attention in which groups of query heads share key and value heads, shrinking the key-value cache and decode memory traffic."@en ; mlsys:reduces ; mlsys:specializes ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Context-length extension"@en ; skos:altLabel "RoPE scaling"@en ; skos:definition "Techniques that extend a trained model's usable sequence length, such as position-encoding interpolation followed by long-sequence fine-tuning."@en ; mlsys:uses ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Weight serialization format"@en ; skos:altLabel "safetensors"@en ; skos:definition "The on-disk encoding of a model's parameter tensors and metadata, governing load time, memory mapping, and safety of deserialization."@en ; mlsys:part-of ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model exchange format"@en ; skos:altLabel "ONNX"@en ; skos:definition "A framework-neutral representation of a model's operator graph and weights used to move models between training frameworks and inference runtimes."@en ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model export"@en ; skos:definition "Converting a framework program into a self-contained deployable artifact by tracing or capturing its graph and fixing operators, shapes, and weights."@en ; mlsys:produces ; mlsys:produces ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model signature"@en ; skos:definition "The declared names, types, and shapes of a model artifact's inputs and outputs, forming its interface contract with callers."@en ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Ridge point"@en ; skos:definition "The arithmetic intensity at which a roofline's bandwidth and compute ceilings meet, equal to peak compute throughput divided by peak memory bandwidth."@en ; mlsys:determines ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Compute-bound vs. memory-bound"@en ; skos:definition "Classifying a kernel or phase by which roofline ceiling limits it, indicating whether more arithmetic throughput or more bandwidth would help."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data-movement energy"@en ; skos:definition "The energy cost of moving operands between storage levels, which grows with distance and typically exceeds the energy of the arithmetic performed on them."@en ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Amdahl's law"@en ; skos:definition "The bound that overall speedup from accelerating part of an execution is limited by the fraction left unaccelerated."@en ; mlsys:uses ; mlsys:limits ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Strong scaling"@en ; skos:definition "How time to solution changes as resources increase for a fixed total problem size."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Weak scaling"@en ; skos:altLabel "Gustafson's law"@en ; skos:definition "How time to solution changes as resources and problem size grow proportionally, keeping work per resource fixed."@en ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Parallel scaling efficiency"@en ; skos:altLabel "scaling efficiency"@en ; skos:definition "Achieved speedup divided by the factor of added resources, measuring how much additional hardware turns into useful progress."@en ; mlsys:measures ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Little's law"@en ; skos:altLabel "L = lambda W"@en ; skos:definition "The relation that the mean number of items in a stable system equals mean arrival rate times mean time in the system."@en ; mlsys:uses ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Latency-throughput curve"@en ; skos:altLabel "load-latency curve"@en ; skos:definition "The relationship between offered load and latency, in which queueing delay grows sharply as utilization approaches capacity."@en ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Arrival burstiness"@en ; skos:definition "Variability of request inter-arrival times or sizes beyond a Poisson baseline, which raises queueing delay at a given mean utilization."@en ; mlsys:increases ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Compute-optimal allocation"@en ; skos:altLabel "Chinchilla scaling"@en ; skos:definition "Choosing model size and training data quantity to minimize loss under a fixed training compute budget, using fitted scaling laws."@en ; mlsys:requires ; mlsys:uses ; mlsys:trades-off . a skos:Concept ; skos:inScheme ; skos:prefLabel "Training compute estimate"@en ; skos:altLabel "6ND"@en ; skos:definition "Approximating dense transformer training compute as roughly six floating-point operations per parameter per training token."@en ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Systolic array"@en ; skos:definition "A grid of processing elements that pass operands rhythmically to neighbors, performing matrix multiplication with high operand reuse and little register-file traffic."@en ; mlsys:implements ; mlsys:specializes ; mlsys:reduces ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Matrix engine"@en ; skos:altLabel "Tensor Core"@en ; skos:altLabel "MXU"@en ; skos:altLabel "tensor core"@en ; skos:definition "A functional unit that executes small dense matrix multiply-accumulate tiles per instruction, supplying most of an accelerator's peak low-precision throughput."@en ; mlsys:requires ; mlsys:determines ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "SIMT execution"@en ; skos:altLabel "warp"@en ; skos:definition "Single-instruction multiple-thread execution, in which groups of threads issue the same instruction and divergent control flow within a group is serialized."@en ; mlsys:determines ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Accelerator dataflow"@en ; skos:altLabel "stationarity"@en ; skos:definition "The mapping that decides which operands stay resident in local storage while others stream, such as weight-, output-, or row-stationary schedules."@en ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Domain-specific architecture"@en ; skos:altLabel "DSA"@en ; skos:definition "Hardware specialized to a class of workloads, trading programmability for higher performance per watt on its target operations."@en ; mlsys:increases ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "High-bandwidth memory"@en ; skos:altLabel "HBM"@en ; skos:definition "Stacked DRAM placed beside the processor on an interposer, providing very high bandwidth with limited capacity per stack."@en ; mlsys:increases ; mlsys:part-of ; mlsys:determines ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "On-chip SRAM and scratchpads"@en ; skos:definition "Small, fast on-die storage, such as caches, shared memory, and register files, used to stage reused operands near compute units."@en ; mlsys:limits ; mlsys:limits ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Memory wall"@en ; skos:definition "The widening gap between processor arithmetic throughput and memory bandwidth or latency, which makes data movement the limiting factor."@en ; mlsys:determines ; mlsys:reduces ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Scale-up interconnect"@en ; skos:definition "High-bandwidth links and switches, such as NVLink-class fabrics, joining accelerators within a node or rack into one tightly coupled domain."@en ; mlsys:limits ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Cache-coherent interconnect"@en ; skos:altLabel "CXL"@en ; skos:definition "A link protocol that lets hosts and devices share or pool memory with hardware coherence, such as CXL over a PCIe physical layer."@en ; mlsys:increases ; mlsys:specializes ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Peer-to-peer device transfer"@en ; skos:altLabel "GPUDirect"@en ; skos:definition "Direct memory access between accelerators, or between accelerators and NICs or storage, without staging data through host memory."@en ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Thermal design power"@en ; skos:altLabel "TDP"@en ; skos:definition "The sustained heat dissipation a cooling solution is designed to remove, used as a chip's nominal power envelope."@en ; mlsys:limits ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Dynamic voltage and frequency scaling"@en ; skos:altLabel "DVFS"@en ; skos:definition "Adjusting supply voltage and clock frequency at runtime to trade performance for power, since dynamic power grows superlinearly with voltage."@en ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Dark silicon"@en ; skos:definition "The fraction of a chip that cannot be powered at full frequency simultaneously within its power and thermal budget."@en ; mlsys:enables ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Power capping"@en ; skos:definition "Enforcing a maximum power draw on a device or group by limiting clocks or activity, trading peak throughput for bounded power."@en ; mlsys:reduces ; mlsys:uses ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Microcontroller-class device"@en ; skos:altLabel "MCU"@en ; skos:altLabel "TinyML device"@en ; skos:definition "A low-power processor with on-chip flash and kilobytes to a few megabytes of SRAM, typically without an operating system or virtual memory."@en ; mlsys:determines ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Heterogeneous mobile SoC"@en ; skos:definition "A system-on-chip combining CPU, GPU, DSP, and NPU blocks under shared memory and power limits, across which model operators are partitioned."@en ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "On-device neural accelerator"@en ; skos:altLabel "NPU"@en ; skos:definition "An accelerator integrated in mobile or embedded silicon for low-precision neural inference, usually supporting a restricted set of operators."@en ; mlsys:requires ; mlsys:part-of ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Energy per inference"@en ; skos:definition "Energy consumed to process one input end to end on a device, the binding metric for battery-powered and energy-harvesting deployments."@en ; mlsys:specializes ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "On-device memory budget"@en ; skos:altLabel "app memory limit"@en ; skos:altLabel "mobile memory budget"@en ; skos:definition "The flash and SRAM limits that bound a deployed model's stored weights and peak activation memory on embedded hardware."@en ; mlsys:requires ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Tensor"@en ; skos:definition "A multidimensional array carrying shape, element type, device, and layout metadata, the core data abstraction of ML frameworks."@en ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Computation graph"@en ; skos:definition "A directed graph of tensor operations and their data dependencies representing a model program, used for differentiation, optimization, and scheduling."@en ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Eager vs. graph execution"@en ; skos:altLabel "define-by-run"@en ; skos:definition "The choice between running each operation as the host program issues it and building a whole graph first for later optimization and execution."@en ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Tracing and JIT capture"@en ; skos:altLabel "JIT compilation"@en ; skos:definition "Recording operations executed on example inputs to build a graph compiled just in time, with guards or retracing when shapes or control flow change."@en ; mlsys:enables ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "IO-aware attention kernel"@en ; skos:altLabel "FlashAttention"@en ; skos:definition "An exact attention implementation that tiles queries, keys, and values through on-chip memory to avoid materializing the full score matrix in off-chip memory."@en ; mlsys:requires ; mlsys:implements ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Convolution lowering"@en ; skos:altLabel "im2col"@en ; skos:definition "Implementing convolution by rearranging input patches into matrices for GEMM, or by direct, FFT, or Winograd algorithms with different memory costs."@en ; mlsys:uses ; mlsys:implements ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Vendor kernel library"@en ; skos:altLabel "cuBLAS"@en ; skos:altLabel "cuDNN"@en ; skos:definition "Tuned implementations of common operators supplied for a hardware platform, which frameworks call rather than generating their own kernels."@en ; mlsys:reduces ; mlsys:implements ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "ML compiler IR"@en ; skos:altLabel "graph IR"@en ; skos:altLabel "MLIR"@en ; skos:definition "Program representations at graph, tensor, and loop levels on which compilers apply lowering and optimization passes."@en ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Kernel autotuning"@en ; skos:definition "Searching kernel schedule parameters, such as tile sizes and loop orders, by measurement or learned cost models to find fast implementations for a target."@en ; mlsys:determines ; mlsys:uses ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Kernel DSL"@en ; skos:altLabel "Triton"@en ; skos:definition "A domain-specific language for writing tiled tensor kernels above the level of vendor assembly, with the compiler handling memory and thread mapping."@en ; mlsys:uses ; mlsys:uses ; mlsys:enables ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Static memory planning"@en ; skos:definition "Compile-time assignment of tensor buffers to memory with lifetime-based reuse, reducing peak memory and runtime allocation."@en ; mlsys:reduces ; mlsys:part-of ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Device streams"@en ; skos:altLabel "CUDA streams"@en ; skos:definition "Ordered queues of device work that the host enqueues asynchronously, with events expressing dependencies and allowing overlap across queues."@en ; mlsys:enables ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Caching memory allocator"@en ; skos:definition "A runtime allocator that retains and reuses freed device memory blocks to avoid costly allocation calls, at the risk of fragmentation."@en ; mlsys:reduces ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Captured graph replay"@en ; skos:altLabel "CUDA Graphs"@en ; skos:definition "Recording a fixed sequence of kernel launches once and replaying it as a unit to remove per-launch host overhead."@en ; mlsys:reduces ; mlsys:uses ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Loss scaling"@en ; skos:definition "Multiplying the loss before backpropagation so small gradients remain representable in reduced-precision formats, then unscaling before the optimizer step."@en ; mlsys:part-of ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Block-scaled number formats"@en ; skos:altLabel "MX"@en ; skos:altLabel "microscaling"@en ; skos:altLabel "MXFP8"@en ; skos:altLabel "FP8 with block scaling"@en ; skos:definition "Low-precision formats in which small blocks of values share a scale factor, extending dynamic range for FP8, FP6, FP4, or INT8 elements."@en ; mlsys:increases ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Accumulation precision"@en ; skos:definition "The numerical format used to sum products in dot products and reductions, which may differ from input precision and governs rounding-error growth."@en ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Expert parallelism"@en ; skos:altLabel "EP"@en ; skos:definition "Placing different experts of a mixture-of-experts layer on different devices and exchanging routed tokens among them."@en ; mlsys:requires ; mlsys:enables ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Sequence and context parallelism"@en ; skos:altLabel "context parallelism"@en ; skos:altLabel "ring attention"@en ; skos:definition "Partitioning activations along the sequence dimension across devices, exchanging keys, values, or partial results so long sequences fit."@en ; mlsys:increases ; mlsys:part-of ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Hybrid parallelism"@en ; skos:altLabel "3D parallelism"@en ; skos:definition "Combining data, tensor, pipeline, and other parallelism dimensions, mapped onto the hardware hierarchy so the most communication-intensive dimension uses the fastest links."@en ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Pipeline bubble"@en ; skos:definition "Idle device time during pipeline fill and drain, whose fraction shrinks as the number of microbatches grows relative to pipeline stages."@en ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Ring all-reduce"@en ; skos:definition "An all-reduce performed as reduce-scatter then all-gather around a ring, bandwidth-optimal for large messages but with latency linear in participants."@en ; mlsys:implements . a skos:Concept ; skos:inScheme ; skos:prefLabel "Tree and hierarchical all-reduce"@en ; skos:definition "All-reduce schedules over trees or hierarchical groups, with latency logarithmic in participants, suited to small messages and multi-level networks."@en ; mlsys:alternative-to ; mlsys:implements . a skos:Concept ; skos:inScheme ; skos:prefLabel "All-to-all exchange"@en ; skos:altLabel "AllToAll"@en ; skos:definition "A collective in which every participant sends distinct data to every other, as in expert routing; its throughput is sensitive to bisection bandwidth."@en ; mlsys:specializes ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Alpha-beta communication model"@en ; skos:altLabel "latency-bandwidth model"@en ; skos:altLabel "network transfer time"@en ; skos:altLabel "transfer time"@en ; skos:definition "A cost model estimating message time as a fixed latency term plus message size times a per-byte cost, used to compare collective algorithms."@en ; mlsys:measures ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Gradient compression"@en ; skos:definition "Reducing gradient communication through quantization, sparsification, or low-rank approximation, often with error feedback to preserve convergence."@en ; mlsys:reduces ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Lease"@en ; skos:definition "A time-bounded grant of authority or cached state that expires unless renewed, letting a system recover from failed holders without explicit release."@en ; mlsys:requires ; mlsys:enables ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Leader election"@en ; skos:definition "Selecting a single coordinator among replicas, typically through consensus and leases, so control decisions have one authoritative source."@en ; mlsys:uses ; mlsys:uses ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Bounded staleness"@en ; skos:altLabel "stale synchronous parallel"@en ; skos:definition "A consistency model in which workers may compute on parameters up to a fixed number of updates old, trading freshness for less waiting."@en ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Federated averaging"@en ; skos:altLabel "FedAvg"@en ; skos:definition "A federated algorithm in which sampled clients run several local optimization steps and a server averages their model updates, weighted by data size."@en ; mlsys:implements ; mlsys:uses ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Statistical heterogeneity"@en ; skos:altLabel "non-IID data"@en ; skos:definition "Differences in data distribution and quantity across federated clients, which slow convergence and can bias a shared model."@en ; mlsys:limits ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Client selection"@en ; skos:definition "Choosing which available devices participate in each federated round, subject to availability, connectivity, and power constraints."@en ; mlsys:increases ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Cross-device vs. cross-silo federation"@en ; skos:definition "The distinction between federations of many intermittently available devices and federations of a few reliable organizational data holders."@en ; mlsys:determines ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Power usage effectiveness"@en ; skos:altLabel "PUE"@en ; skos:definition "Total facility energy divided by IT equipment energy, measuring overhead from cooling, power conversion, and other non-compute loads."@en ; mlsys:determines ; mlsys:measures ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Liquid cooling"@en ; skos:definition "Removing heat with liquid through cold plates or immersion, enabling rack power densities beyond the practical limits of air cooling."@en ; mlsys:enables ; mlsys:reduces ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Rack power density"@en ; skos:definition "Electrical power drawn per rack, which determines cooling technology, floor layout, and how many accelerators fit within a facility's power envelope."@en ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Power oversubscription"@en ; skos:definition "Deploying more equipment than the sum of nameplate power a feed supports, relying on non-coincident peaks and power capping to stay within limits."@en ; mlsys:requires ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Warehouse-scale computer"@en ; skos:altLabel "WSC"@en ; skos:definition "A datacenter treated as a single computer whose power, cooling, network, and software are co-designed for a common workload."@en ; mlsys:uses ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Remote direct memory access"@en ; skos:altLabel "RDMA"@en ; skos:altLabel "InfiniBand"@en ; skos:altLabel "RoCE"@en ; skos:definition "Network transfers that read or write remote memory directly through the network interface without remote CPU involvement, giving low latency and high throughput."@en ; mlsys:reduces ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Lossless network fabric"@en ; skos:altLabel "PFC"@en ; skos:definition "A network that avoids packet drops through credit-based or priority flow control, as RDMA transports expect, at the risk of head-of-line blocking."@en ; mlsys:enables ; mlsys:requires ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Datacenter congestion control"@en ; skos:altLabel "DCQCN"@en ; skos:definition "End-host and switch mechanisms, such as ECN-driven rate control, that limit queue buildup from incast and synchronized collective traffic."@en ; mlsys:mitigates ; mlsys:mitigates ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Fat-tree (Clos) topology"@en ; skos:altLabel "Clos"@en ; skos:definition "A multistage switch topology providing many equal-cost paths and, without oversubscription, full bisection bandwidth between endpoints."@en ; mlsys:increases ; mlsys:specializes ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Rail-optimized topology"@en ; skos:definition "A fabric design connecting the same-index accelerator of every node to a common leaf switch, keeping most collective traffic within one switch hop."@en ; mlsys:enables ; mlsys:specializes ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Bisection bandwidth"@en ; skos:definition "The minimum total bandwidth across any cut dividing a network into two equal halves, bounding throughput of all-to-all and global exchanges."@en ; mlsys:limits ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "In-network reduction"@en ; skos:altLabel "SHARP"@en ; skos:definition "Performing collective reductions inside network switches so data crosses each link fewer times, lowering all-reduce traffic and latency."@en ; mlsys:implements ; mlsys:reduces ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Parallel file system"@en ; skos:altLabel "Lustre"@en ; skos:altLabel "GPFS"@en ; skos:definition "A distributed file system that stripes data across many storage servers to deliver aggregate bandwidth to many concurrent clients."@en ; mlsys:enables ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Object storage"@en ; skos:definition "Storage of immutable named objects behind an HTTP-style interface, offering scalable capacity and throughput with higher per-request latency than file systems."@en ; mlsys:part-of ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Training input pipeline"@en ; skos:altLabel "data loader"@en ; skos:definition "The read, decode, transform, shuffle, and prefetch stages that deliver batches to accelerators; when slower than compute, they cause data stalls."@en ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Checkpoint I/O"@en ; skos:definition "Writing and reading large training-state snapshots, where storage bandwidth, asynchronous persistence, and sharded layout determine checkpoint cost."@en ; mlsys:determines ; mlsys:uses ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Storage tiering and caching"@en ; skos:definition "Placing data across tiers, from memory and local NVMe caches to shared and archival storage, according to access frequency and cost."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Gang scheduling"@en ; skos:definition "Allocating all resources a parallel job needs at once, or none, so tightly coupled workers start together and partial allocations cannot deadlock."@en ; mlsys:reduces ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Job preemption"@en ; skos:definition "Suspending or evicting a running job to reclaim its resources for higher-priority work, at a cost set by checkpointing and lost progress."@en ; mlsys:requires ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Dominant resource fairness"@en ; skos:altLabel "DRF"@en ; skos:definition "A multi-resource fair-sharing policy that equalizes each user's largest share across resource types such as accelerators, CPUs, and memory."@en ; mlsys:implements . a skos:Concept ; skos:inScheme ; skos:prefLabel "Elastic training"@en ; skos:definition "Training that continues while the number of workers changes, resharding state and adjusting batch size or learning rate as resources join or leave."@en ; mlsys:mitigates ; mlsys:uses ; mlsys:increases ; mlsys:requires ; mlsys:enables ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Topology-aware placement"@en ; skos:definition "Assigning a job's workers to devices and nodes so that heavily communicating groups share fast links or switches."@en ; mlsys:reduces ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Bin packing"@en ; skos:definition "Placing jobs onto nodes to minimize fragmentation of resources such as accelerators, trading packing density against interference and locality."@en ; mlsys:trades-off ; mlsys:increases ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Multi-tenancy"@en ; skos:definition "Sharing infrastructure among teams, jobs, or customers under policies for allocation, isolation, priority, and accounting."@en ; mlsys:requires ; mlsys:uses ; mlsys:increases ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Resource quota"@en ; skos:definition "A limit on the resources a tenant may hold or request, sometimes with guaranteed minimums and borrowing of idle capacity."@en ; mlsys:part-of ; mlsys:part-of ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Accelerator sharing and partitioning"@en ; skos:altLabel "MIG"@en ; skos:altLabel "GPU sharing"@en ; skos:definition "Running several workloads on one accelerator by time-slicing, concurrent contexts, or hardware partitions with dedicated memory and compute slices."@en ; mlsys:increases ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Spot and preemptible capacity"@en ; skos:definition "Discounted compute that the provider may reclaim at short notice, usable by workloads that tolerate interruption."@en ; mlsys:causes ; mlsys:reduces ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Performance interference"@en ; skos:altLabel "noisy neighbor"@en ; skos:definition "Slowdown of one workload caused by co-located workloads contending for shared caches, memory bandwidth, network, or power."@en ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Weak supervision"@en ; skos:definition "Producing training labels from noisy programmatic sources such as heuristics, rules, or distant supervision, then modeling and combining their errors."@en ; mlsys:determines ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Active learning"@en ; skos:definition "Selecting which unlabeled examples to annotate next according to their expected value to the model."@en ; mlsys:determines ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Inter-annotator agreement"@en ; skos:definition "The degree to which independent annotators assign the same label to the same item, used to measure label reliability and task ambiguity."@en ; mlsys:measures ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Quality filtering"@en ; skos:definition "Removing or down-weighting training examples judged low quality by heuristics, classifiers, or model scores."@en ; mlsys:implements ; mlsys:increases ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data mixture"@en ; skos:altLabel "domain mixture"@en ; skos:definition "The proportions in which source domains or datasets are sampled during training."@en ; mlsys:determines ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Synthetic data"@en ; skos:definition "Training or evaluation examples generated by simulators, programs, or models rather than collected from the target environment."@en ; mlsys:causes ; mlsys:enables ; mlsys:alternative-to ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data augmentation"@en ; skos:definition "Creating additional training examples by applying label-preserving transformations to existing ones."@en ; mlsys:increases ; mlsys:specializes ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data validation"@en ; skos:definition "Automated checks of incoming data against expected schema, value ranges, and distributions before it is used for training or serving."@en ; mlsys:implements ; mlsys:detects ; mlsys:measures ; mlsys:detects ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Feature store"@en ; skos:definition "A system that defines, computes, stores, and serves features consistently for offline training and online inference."@en ; mlsys:mitigates ; mlsys:enables ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Stream processing"@en ; skos:definition "Continuously transforming unbounded event streams with low latency and incrementally maintained state."@en ; mlsys:limits ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Feature backfill"@en ; skos:definition "Recomputing a feature's historical values under a new or corrected definition so training data matches what serving will produce."@en ; mlsys:requires ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Dataset versioning"@en ; skos:definition "Recording immutable, addressable versions of datasets so that a training run or evaluation can identify exactly which data it used."@en ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data dependency debt"@en ; skos:definition "Undeclared, unstable, or underused input signals on which a model depends, which can change or break its behavior silently."@en ; mlsys:increases ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Feature freshness"@en ; skos:definition "The age of a feature value at the time it is used for a prediction, relative to the freshness assumed during training."@en ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Learning-rate schedule"@en ; skos:definition "The planned variation of the optimizer step size over training, such as warmup followed by decay."@en ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Critical batch size"@en ; skos:definition "The batch size beyond which increasing it further yields diminishing reductions in optimization steps to reach a target loss."@en ; mlsys:limits ; mlsys:limits ; mlsys:limits ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Training instability"@en ; skos:definition "Loss spikes, divergence, or collapse during training, often mitigated by gradient clipping, schedule changes, normalization, or restarting from a checkpoint."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Gradient accumulation"@en ; skos:definition "Summing gradients over several micro-batches before one optimizer update to emulate a larger batch within limited memory."@en ; mlsys:enables ; mlsys:reduces ; mlsys:reduces ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Training-state offloading"@en ; skos:definition "Moving optimizer state, parameters, or activations from accelerator memory to host memory or storage between uses."@en ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Experiment tracking"@en ; skos:definition "Recording each run's code, data, configuration, metrics, and artifacts in a queryable store for comparison and reproduction."@en ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Hyperparameter transfer"@en ; skos:altLabel "muP"@en ; skos:definition "Tuning hyperparameters on smaller proxy models and reusing them at larger scale under a parameterization designed to keep them stable."@en ; mlsys:reduces ; mlsys:determines ; mlsys:uses ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Supervised fine-tuning"@en ; skos:altLabel "SFT"@en ; skos:altLabel "instruction tuning"@en ; skos:definition "Continued training of a pretrained model on input-output demonstrations, such as instruction-response pairs, to shape task behavior."@en ; mlsys:causes ; mlsys:specializes ; mlsys:implements . a skos:Concept ; skos:inScheme ; skos:prefLabel "Low-rank adaptation"@en ; skos:altLabel "LoRA"@en ; skos:definition "Fine-tuning by learning low-rank update matrices added to frozen pretrained weights."@en ; mlsys:specializes ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "RL from human feedback"@en ; skos:altLabel "RLHF"@en ; skos:definition "Optimizing a model's policy against a reward model trained on human preference comparisons, usually with a penalty for drifting from a reference model."@en ; mlsys:causes ; mlsys:requires ; mlsys:specializes ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Reward model"@en ; skos:definition "A learned model that scores outputs or trajectories to supply the training signal for policy optimization."@en ; mlsys:part-of ; mlsys:causes ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Direct preference optimization"@en ; skos:altLabel "DPO"@en ; skos:definition "Training a policy directly on preferred-versus-rejected output pairs with a classification-style loss, without a separate reward model or RL loop."@en ; mlsys:alternative-to ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "RL with verifiable rewards"@en ; skos:altLabel "RLVR"@en ; skos:altLabel "GRPO-style training"@en ; skos:definition "Reinforcement learning in which rewards come from programmatic checks of outcomes, such as answer matching or test execution, rather than learned preference models."@en ; mlsys:requires ; mlsys:specializes ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "RL rollout generation"@en ; skos:definition "Sampling responses or trajectories from the current policy, often with an inference engine, to produce experience for policy updates."@en ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Policy weight synchronization"@en ; skos:definition "Transferring updated policy weights from training workers to rollout workers, whose cost and frequency shape RL post-training throughput."@en ; mlsys:determines ; mlsys:uses ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Policy staleness"@en ; skos:altLabel "off-policy lag"@en ; skos:definition "The lag between the policy version that generated rollouts and the version being updated, introduced by asynchronous rollout and training."@en ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Online learning"@en ; skos:definition "Updating a model incrementally as each example or small batch arrives, rather than retraining on a fixed dataset."@en ; mlsys:determines ; mlsys:uses ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Incremental training"@en ; skos:altLabel "stateful training"@en ; skos:altLabel "warm-start training"@en ; skos:definition "Continuing training of an existing model from its current weights on new data instead of training from scratch."@en ; mlsys:specializes ; mlsys:causes ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Catastrophic forgetting"@en ; skos:definition "Loss of previously learned capability when a model is trained on new data or tasks."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Experience replay"@en ; skos:definition "Mixing stored past examples into updates on new data to reduce forgetting or reuse experience."@en ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "On-device training"@en ; skos:definition "Updating model parameters on the deployment device using local data under its memory, energy, and compute limits."@en ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Cross-validation"@en ; skos:definition "Estimating performance by repeatedly training on some data partitions and evaluating on the held-out remainder."@en ; mlsys:reduces ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Sequential testing"@en ; skos:altLabel "SPRT"@en ; skos:altLabel "always-valid inference"@en ; skos:definition "Experimental analysis that permits repeated looks at accumulating data while controlling error rates, unlike fixed-horizon tests."@en ; mlsys:enables ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Experiment interference"@en ; skos:altLabel "SUTVA violation"@en ; skos:definition "Treatment of one experimental unit affecting outcomes of another, violating the no-interference assumption (SUTVA) of standard A/B analysis."@en ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Conformal prediction"@en ; skos:definition "Constructing prediction sets with finite-sample coverage guarantees under exchangeability, using held-out calibration scores."@en ; mlsys:enables ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Slice discovery"@en ; skos:definition "Automatically finding coherent data subsets on which a model underperforms, without predefined slice labels."@en ; mlsys:enables ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model-based grading"@en ; skos:altLabel "LLM-as-judge"@en ; skos:definition "Using a language model to score or compare generated outputs against a rubric, with its own biases and error rates to be measured."@en ; mlsys:requires ; mlsys:detects . a skos:Concept ; skos:inScheme ; skos:prefLabel "pass@k"@en ; skos:definition "The probability that at least one of k sampled outputs satisfies a correctness check."@en ; mlsys:limits ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Pairwise preference evaluation"@en ; skos:definition "Ranking models from judgments of which of two outputs is better, typically aggregated with a Bradley-Terry or Elo-style model."@en ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Human evaluation"@en ; skos:definition "Assessment of system outputs by people following a protocol, used where automatic metrics or model graders are not trusted."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Benchmark contamination"@en ; skos:definition "Presence of evaluation items, or close variants, in a model's training data, inflating measured performance."@en ; mlsys:specializes ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Time to train"@en ; skos:definition "Wall-clock time for a system to train a specified model to a specified quality target under fixed run rules."@en ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Inference load scenario"@en ; skos:definition "A specified query-arrival pattern and latency constraint, such as single-stream, server, or offline, under which inference performance is measured."@en ; mlsys:limits ; mlsys:determines ; mlsys:uses ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Metamorphic testing"@en ; skos:definition "Testing by checking that known relations hold between outputs for related inputs, when correct outputs for single inputs are unknown."@en ; mlsys:implements . a skos:Concept ; skos:inScheme ; skos:prefLabel "Behavioral testing"@en ; skos:definition "Testing model capabilities with targeted input suites for invariance, directional change, and minimum functionality, independent of aggregate accuracy."@en ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Production-readiness assessment"@en ; skos:definition "A structured rubric of data, model, infrastructure, and monitoring tests used to judge whether an ML system is ready to operate."@en ; mlsys:determines ; mlsys:measures ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Online inference"@en ; skos:altLabel "real-time prediction"@en ; skos:definition "Producing predictions on demand for individual requests within a latency target."@en ; mlsys:alternative-to ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Batch inference"@en ; skos:altLabel "offline prediction"@en ; skos:definition "Computing predictions for many inputs in scheduled jobs and storing them for later lookup."@en ; mlsys:reduces ; mlsys:reduces ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model server"@en ; skos:altLabel "inference server"@en ; skos:definition "A service that loads model artifacts and executes inference requests behind a network interface, managing versions, batching, and resources."@en ; mlsys:uses ; mlsys:requires ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Multi-model serving"@en ; skos:definition "Hosting many models or adapters on shared serving resources, trading isolation and load time against utilization."@en ; mlsys:increases ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model cascade"@en ; skos:definition "Sending a request to cheaper models first and escalating to more capable models only when a confidence or quality criterion is not met."@en ; mlsys:reduces ; mlsys:requires ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model routing"@en ; skos:definition "Choosing which model serves each request from request features, predicted quality, and cost."@en ; mlsys:requires ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Time to first token"@en ; skos:altLabel "TTFT"@en ; skos:definition "Latency from request arrival until the first generated token is returned, dominated by queueing and prefill."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Time per output token"@en ; skos:altLabel "TPOT"@en ; skos:altLabel "inter-token latency"@en ; skos:definition "Average latency between successive generated tokens after the first, dominated by decode step time."@en ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Hedged requests"@en ; skos:definition "Issuing a duplicate request to another replica when the first has not responded within a delay, and using whichever returns first."@en ; mlsys:reduces ; mlsys:increases ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Continuous batching"@en ; skos:altLabel "iteration-level batching"@en ; skos:altLabel "in-flight batching"@en ; skos:definition "Scheduling at each generation iteration so finished sequences leave and new requests join a running batch."@en ; mlsys:specializes ; mlsys:increases ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Request load balancing"@en ; skos:definition "Distributing requests across replicas according to load, locality, or cached state."@en ; mlsys:determines ; mlsys:reduces ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model cold start"@en ; skos:definition "Delay before a newly started replica can serve, from provisioning, loading weights, and warming runtime state."@en ; mlsys:limits ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Chunked prefill"@en ; skos:definition "Splitting a long prompt's prefill into chunks interleaved with decode steps to bound per-iteration latency."@en ; mlsys:reduces ; mlsys:alternative-to ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Prefill-decode disaggregation"@en ; skos:altLabel "disaggregated serving"@en ; skos:definition "Running prefill and decode on separate resource pools, transferring KV state between them, so each phase is provisioned and batched independently."@en ; mlsys:determines ; mlsys:increases ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Decoding and sampling controls"@en ; skos:definition "Parameters and rules, such as greedy, temperature, top-k, and nucleus sampling, that determine how tokens are chosen from a model's distribution."@en ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Constrained decoding"@en ; skos:altLabel "grammar-guided decoding"@en ; skos:definition "Masking disallowed tokens at each step, typically from a grammar or schema automaton, so output satisfies a structural constraint."@en ; mlsys:implements ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Output validation"@en ; skos:definition "Checking generated output against a schema or rules after generation and repairing, retrying, or rejecting it on failure."@en ; mlsys:alternative-to ; mlsys:implements ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Reasoning budget"@en ; skos:altLabel "thinking budget"@en ; skos:definition "A limit on the intermediate tokens or steps a model may spend before answering, set to trade quality against latency and cost."@en ; mlsys:increases ; mlsys:limits ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Best-of-n sampling"@en ; skos:definition "Generating n candidate outputs and returning the one ranked highest by a scorer or verifier."@en ; mlsys:increases ; mlsys:alternative-to ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Self-consistency voting"@en ; skos:altLabel "majority voting"@en ; skos:definition "Sampling multiple reasoning paths and returning the most frequent final answer."@en ; mlsys:requires ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Verifier"@en ; skos:altLabel "outcome reward model"@en ; skos:definition "A model or program that judges whether a candidate output or reasoning step is correct, used for selection, search, or reward."@en ; mlsys:limits ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Process reward model"@en ; skos:altLabel "PRM"@en ; skos:definition "A verifier that scores intermediate reasoning steps rather than only final answers."@en ; mlsys:specializes ; mlsys:specializes ; mlsys:implements . a skos:Concept ; skos:inScheme ; skos:prefLabel "Inference-time search"@en ; skos:definition "Exploring a tree of partial solutions, as in beam or tree search, guided by a scorer, to find a higher-quality output."@en ; mlsys:uses ; mlsys:increases ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Test-time scaling"@en ; skos:definition "The empirical relation between inference compute spent per query and output quality, used to allocate compute between model size and inference."@en ; mlsys:specializes ; mlsys:determines ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Retrieval-augmented generation"@en ; skos:altLabel "RAG"@en ; skos:definition "Conditioning generation on documents retrieved for the request so outputs can draw on information outside the model's parameters."@en ; mlsys:mitigates ; mlsys:requires ; mlsys:specializes ; mlsys:increases ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Dense retrieval"@en ; skos:definition "Retrieving records by similarity between learned embeddings of query and records."@en ; mlsys:uses ; mlsys:uses ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Lexical retrieval"@en ; skos:altLabel "sparse retrieval"@en ; skos:altLabel "BM25"@en ; skos:definition "Retrieving records by weighted term overlap with the query using an inverted index, as in BM25."@en ; mlsys:alternative-to ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Reranking"@en ; skos:definition "Reordering a retrieved candidate set with a more expensive relevance model before use."@en ; mlsys:part-of ; mlsys:mitigates ; mlsys:part-of ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model conversion"@en ; skos:altLabel "model export"@en ; skos:definition "Translating a trained model into the format and operator set of a target runtime, possibly changing numerics and supported operations."@en ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "On-device inference runtime"@en ; skos:definition "An interpreter or compiled engine that executes models on mobile, embedded, or microcontroller targets within their memory and operator limits."@en ; mlsys:uses ; mlsys:requires ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Over-the-air model update"@en ; skos:altLabel "OTA update"@en ; skos:definition "Delivering new model versions to deployed devices remotely, with staging, verification, and rollback across the device fleet."@en ; mlsys:uses ; mlsys:requires ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Device-cloud partitioning"@en ; skos:altLabel "split computing"@en ; skos:altLabel "offloading"@en ; skos:definition "Splitting inference between device and server, by layer or by model, to balance latency, energy, bandwidth, and privacy."@en ; mlsys:specializes ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Device heterogeneity"@en ; skos:definition "Variation in hardware, operating system, and runtime capability across a deployed device fleet that a single model release must accommodate."@en ; mlsys:increases ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model registry"@en ; skos:definition "A versioned store of model artifacts with metadata, lineage, and lifecycle stage used to govern promotion and deployment."@en ; mlsys:implements ; mlsys:part-of ; mlsys:enables ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Shadow deployment"@en ; skos:altLabel "shadow mode"@en ; skos:definition "Running a candidate model on live traffic without serving its outputs, to compare it against the production model."@en ; mlsys:part-of ; mlsys:detects ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "CI/CD for ML"@en ; skos:definition "Automated building, testing, and deployment of ML code, data checks, and model artifacts on each change."@en ; mlsys:implements ; mlsys:uses ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Release gate"@en ; skos:altLabel "validation gate"@en ; skos:definition "A predeclared set of checks a candidate model must pass before it advances to a wider deployment stage."@en ; mlsys:determines ; mlsys:requires ; mlsys:uses ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Drift detection"@en ; skos:definition "Statistical tests or distances comparing production and reference distributions of inputs, predictions, or outcomes to flag change."@en ; mlsys:detects ; mlsys:part-of ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Prediction logging"@en ; skos:definition "Recording requests, model versions, outputs, and later outcomes so production behavior can be analyzed, debugged, and reused as data."@en ; mlsys:causes ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Alerting"@en ; skos:definition "Rules that turn monitored signals into notifications for responders, tuned to balance missed incidents against alert fatigue."@en ; mlsys:part-of ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Concept drift"@en ; skos:definition "Change over time in the relationship between inputs and targets, so the same input warrants a different prediction."@en ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Continuous training"@en ; skos:altLabel "CT"@en ; skos:definition "An automated pipeline that retrains, validates, and proposes models on a schedule or when triggered by data or performance signals."@en ; mlsys:implements ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Blameless postmortem"@en ; skos:definition "A written analysis after an incident of its timeline, causes, and corrective actions, focused on system fixes rather than blame."@en ; mlsys:part-of ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "ML failure attribution"@en ; skos:definition "Determining whether a production failure stems from data, model, code, configuration, or infrastructure, or from an adversary."@en ; mlsys:requires ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "ML configuration management"@en ; skos:definition "Treating model, feature, and pipeline configuration as versioned, reviewed, and tested code."@en ; mlsys:reduces ; mlsys:enables ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Infrastructure as code"@en ; skos:altLabel "IaC"@en ; skos:definition "Declaring compute, storage, and service infrastructure in versioned machine-readable files from which environments are provisioned."@en ; mlsys:determines ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "ML platform"@en ; skos:definition "Shared internal infrastructure and services that standardize how teams train, deploy, and operate models."@en ; mlsys:reduces ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "ML technical debt"@en ; skos:definition "Accumulated system costs from shortcuts in ML code, data dependencies, configuration, and pipelines that make future change slower and riskier."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Entanglement"@en ; skos:altLabel "CACE"@en ; skos:definition "The property that changing any input, feature, or hyperparameter of a model can change its behavior on everything else."@en ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Undeclared consumers"@en ; skos:definition "Systems that use a model's outputs without the model owner's knowledge, creating hidden coupling."@en ; mlsys:causes ; mlsys:limits ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "MLOps maturity"@en ; skos:definition "Staged levels of automation, from manual training and deployment to automated continuous training and delivery."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "ML team roles"@en ; skos:definition "The division of responsibilities among data, modeling, platform, and operations roles that owns an ML system across its lifecycle."@en ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model deprecation"@en ; skos:definition "Planned withdrawal of a model from service, including notifying and migrating its consumers before removing it."@en ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Artifact archival"@en ; skos:definition "Preserving retired models, data references, and evaluation records needed for audit, reproduction, or rollback after service ends."@en ; mlsys:enables ; mlsys:part-of ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data retention"@en ; skos:definition "Policy and mechanisms that set how long data, logs, and derived artifacts are kept and when they are deleted."@en ; mlsys:implements ; mlsys:trades-off ; mlsys:limits ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model FLOPs utilization"@en ; skos:altLabel "MFU"@en ; skos:definition "Model-required floating-point operations per second achieved, divided by the aggregate peak FLOP rate of the hardware used, excluding recomputation."@en ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Bottleneck analysis"@en ; skos:definition "Identifying the resource or stage, such as compute, memory bandwidth, I/O, or communication, whose capacity limits end-to-end throughput or latency for a workload."@en ; mlsys:uses ; mlsys:detects . a skos:Concept ; skos:inScheme ; skos:prefLabel "Training goodput"@en ; skos:altLabel "ML goodput"@en ; skos:definition "The fraction of allocated accelerator time that produces retained training progress, after losses to scheduling, failures, restarts, and lost work."@en ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Pruning"@en ; skos:definition "Removing weights, neurons, channels, or other structures judged unimportant from a network, usually followed by fine-tuning to recover quality."@en ; mlsys:produces ; mlsys:reduces ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Structured pruning"@en ; skos:definition "Pruning whole channels, filters, heads, or layers so that the remaining model stays dense and runs faster without sparse-execution support."@en ; mlsys:specializes ; mlsys:alternative-to ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Low-rank factorization"@en ; skos:altLabel "low-rank approximation"@en ; skos:definition "Approximating a weight matrix or tensor by a product of smaller factors of reduced rank to cut parameters and operations."@en ; mlsys:reduces ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Quantization-aware training"@en ; skos:altLabel "QAT"@en ; skos:definition "Training or fine-tuning with simulated quantization in the forward pass so the model adapts to reduced precision, in contrast to post-training quantization."@en ; mlsys:specializes ; mlsys:alternative-to ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Neural architecture search"@en ; skos:altLabel "NAS"@en ; skos:definition "Automated search over a defined space of network architectures for designs that optimize a stated objective such as accuracy under a cost budget."@en ; mlsys:specializes ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Hardware-aware NAS"@en ; skos:altLabel "platform-aware NAS"@en ; skos:definition "Architecture search whose objective includes latency, energy, or memory measured or predicted on the target device rather than proxy counts such as FLOPs."@en ; mlsys:specializes ; mlsys:reduces ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Depthwise separable convolution"@en ; skos:definition "Factoring a standard convolution into a per-channel spatial convolution followed by a pointwise convolution, reducing operations and parameters."@en ; mlsys:reduces ; mlsys:specializes ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Accuracy-efficiency frontier"@en ; skos:altLabel "Pareto frontier"@en ; skos:definition "The set of model configurations for which no alternative improves quality without worsening a stated cost measure on a stated platform."@en ; mlsys:requires ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model-hardware co-design"@en ; skos:definition "Jointly choosing model architecture and hardware or kernel design so that each exploits the other's structure."@en ; mlsys:determines ; mlsys:determines ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Retry with backoff"@en ; skos:definition "Reissuing a failed request after increasing, typically jittered, delays and within a retry budget, so that transient faults recover without overloading dependencies."@en ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Circuit breaker"@en ; skos:definition "A client-side guard that stops sending requests to a failing dependency after an error threshold and probes it before resuming traffic."@en ; mlsys:limits ; mlsys:enables ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Replication and failover"@en ; skos:definition "Maintaining redundant replicas of a service or state and redirecting work to a healthy replica when one fails."@en ; mlsys:increases ; mlsys:requires ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Graceful degradation"@en ; skos:definition "Continuing to serve with reduced quality, features, or freshness under failure or overload instead of failing outright."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Optimal checkpoint interval"@en ; skos:altLabel "Young-Daly interval"@en ; skos:definition "The checkpoint period that minimizes expected lost time, balancing checkpoint write cost against expected rework, approximately the square root of twice the write cost times MTBF."@en ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Asynchronous checkpointing"@en ; skos:definition "Snapshotting training state quickly to host memory and persisting it in the background so computation resumes before the durable write completes."@en ; mlsys:reduces ; mlsys:implements ; mlsys:requires ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Silent data corruption"@en ; skos:altLabel "SDC"@en ; skos:definition "Hardware producing wrong computational results without raising an error or crash, so the corruption propagates undetected."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Mean time between failures"@en ; skos:altLabel "MTBF"@en ; skos:definition "The expected operating time between failures of a component or system; for independent components, system MTBF shrinks roughly in inverse proportion to component count."@en ; mlsys:determines ; mlsys:increases ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Availability"@en ; skos:definition "The fraction of time or requests for which a system performs its specified function, approximately MTBF divided by MTBF plus mean time to repair."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Gray failure"@en ; skos:definition "A partial or degraded failure that some observers detect and others do not, such as a slow or lossy node that still passes health checks."@en ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Failure domain"@en ; skos:altLabel "correlated failure"@en ; skos:definition "A set of components that can fail together from a shared cause, such as a rack, power feed, switch, or software release."@en ; mlsys:limits ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Algorithm-based fault tolerance"@en ; skos:altLabel "ABFT"@en ; skos:definition "Detecting or correcting computational errors by carrying checksums through an algorithm, such as matrix multiplication, and verifying invariants on its outputs."@en ; mlsys:detects ; mlsys:uses ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Chaos engineering"@en ; skos:definition "Running controlled experiments that inject real-world failures into production-like systems to test a hypothesis about steady-state behavior."@en ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Game day"@en ; skos:altLabel "disaster recovery exercise"@en ; skos:definition "A scheduled exercise in which teams deliberately trigger failure scenarios to rehearse detection, response, and recovery procedures."@en ; mlsys:measures ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Hardware fault injection"@en ; skos:definition "Emulating hardware faults such as bit flips in weights, activations, or registers to measure how errors propagate to model outputs."@en ; mlsys:measures ; mlsys:specializes ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Blast radius"@en ; skos:definition "The extent of users, components, or effects that a single fault, change, or experiment can impact."@en ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Backdoor attack"@en ; skos:altLabel "trojan attack"@en ; skos:definition "Implanting a hidden behavior in a model, via poisoned data or modified weights, that activates only on inputs containing an attacker-chosen trigger."@en ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Jailbreak"@en ; skos:definition "A user-crafted input that induces a model to bypass its safety training or usage policy and produce disallowed output."@en ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Adversarial training"@en ; skos:definition "Training on adversarially perturbed inputs generated during training to increase robustness within a specified perturbation set."@en ; mlsys:mitigates ; mlsys:trades-off ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model extraction"@en ; skos:altLabel "model stealing"@en ; skos:definition "Reconstructing a functionally equivalent copy or the parameters of a model by querying its prediction interface."@en ; mlsys:produces ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model watermarking"@en ; skos:altLabel "model fingerprinting"@en ; skos:definition "Embedding a verifiable signature in a model's parameters or behavior so that ownership of copies or derivatives can later be demonstrated."@en ; mlsys:detects . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model artifact signing"@en ; skos:definition "Cryptographically signing model files and their provenance metadata so that consumers can verify integrity and origin before loading."@en ; mlsys:measures ; mlsys:uses ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Unsafe model deserialization"@en ; skos:definition "Code execution triggered by loading a model stored in an executable serialization format, such as pickle, from an untrusted source."@en ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "ML supply-chain compromise"@en ; skos:definition "Introducing malicious or vulnerable components into an ML system through third-party datasets, pretrained weights, libraries, build tools, or hardware."@en ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Membership inference"@en ; skos:definition "Inferring whether a specific record was in a model's training data from the model's outputs or parameters."@en ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model inversion"@en ; skos:definition "Reconstructing sensitive attributes or representative training inputs from a model's outputs, confidence scores, or gradients."@en ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "DP-SGD"@en ; skos:altLabel "differentially private SGD"@en ; skos:definition "Stochastic gradient descent with per-example gradient clipping and calibrated noise, with privacy loss tracked across steps by an accountant."@en ; mlsys:implements ; mlsys:reduces ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Machine unlearning"@en ; skos:definition "Removing the influence of specified training records from a trained model, either exactly by retraining or approximately at lower cost."@en ; mlsys:alternative-to ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Secure aggregation"@en ; skos:definition "A cryptographic protocol that lets a server compute the sum of client updates without seeing any individual client's update."@en ; mlsys:increases ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Least privilege"@en ; skos:definition "Granting each component or actor only the permissions its task requires, for only as long as it requires them."@en ; mlsys:limits ; mlsys:mitigates ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Confidential computing"@en ; skos:altLabel "trusted execution environment"@en ; skos:altLabel "TEE"@en ; skos:definition "Protecting data and models while in use by running computation inside hardware-isolated, attestable trusted execution environments."@en ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Remote attestation"@en ; skos:definition "A procedure by which a platform proves to a remote verifier the identity and measured state of its hardware and software before receiving secrets."@en ; mlsys:enables ; mlsys:detects . a skos:Concept ; skos:inScheme ; skos:prefLabel "Side-channel attack"@en ; skos:definition "Inferring secrets from observable resource behavior, such as timing, power, memory, or cache contention, across an isolation boundary."@en ; mlsys:limits ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Energy efficiency"@en ; skos:altLabel "performance per watt"@en ; skos:definition "Useful work completed per unit energy, such as inferences or tokens per joule, measured within a stated system boundary."@en ; mlsys:reduces ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Grid carbon intensity"@en ; skos:definition "Greenhouse-gas emissions per unit of electricity supplied, varying by region, time, and generation mix, in grams CO2-equivalent per kilowatt-hour."@en ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Operational carbon"@en ; skos:definition "Emissions from electricity consumed while operating a system, computed from energy use and the carbon intensity of its supply."@en ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Emission scopes"@en ; skos:altLabel "Scope 1, 2, and 3 emissions"@en ; skos:altLabel "GHG scopes"@en ; skos:definition "The GHG Protocol classification of emissions into direct (Scope 1), purchased energy (Scope 2), and value-chain (Scope 3) categories."@en ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Carbon-aware scheduling"@en ; skos:definition "Shifting flexible workloads in time or location toward periods and regions of lower grid carbon intensity, subject to deadlines and capacity."@en ; mlsys:uses ; mlsys:trades-off . a skos:Concept ; skos:inScheme ; skos:prefLabel "Embodied carbon"@en ; skos:definition "Emissions from manufacturing, transporting, and disposing of hardware and facilities, attributed to a system independent of its operation."@en ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Life cycle assessment"@en ; skos:altLabel "LCA"@en ; skos:definition "A standardized method for compiling and evaluating a product's environmental inputs, outputs, and impacts from raw materials through disposal."@en ; mlsys:measures ; mlsys:measures ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Water footprint"@en ; skos:altLabel "WUE"@en ; skos:definition "Water withdrawn or consumed by a system, on site for cooling and off site for electricity generation, per unit of work or over its lifetime."@en ; mlsys:trades-off . a skos:Concept ; skos:inScheme ; skos:prefLabel "Hardware service lifetime"@en ; skos:altLabel "refresh cycle"@en ; skos:definition "The period a hardware asset stays in service before replacement, over which its embodied impacts and capital cost are amortized."@en ; mlsys:reduces ; mlsys:trades-off ; mlsys:reduces ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Electronic waste"@en ; skos:altLabel "e-waste"@en ; skos:definition "Discarded electronic equipment and components, including retired accelerators and servers, and the hazards and lost materials of their disposal."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Total cost of ownership"@en ; skos:altLabel "TCO"@en ; skos:definition "The full cost of acquiring and operating a system over its life, including capital, energy, facilities, networking, staff, and amortization."@en ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Cost per token"@en ; skos:definition "Serving or API expenditure divided by tokens processed or generated, often priced separately for input and output tokens."@en ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Cost attribution"@en ; skos:altLabel "FinOps"@en ; skos:altLabel "chargeback"@en ; skos:definition "Allocating shared infrastructure spend to teams, models, or workloads so that owners see and act on their costs."@en ; mlsys:enables ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Hazard analysis"@en ; skos:definition "Systematically identifying conditions under which a system could contribute to harm, with their causes, severity, and required controls."@en ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "System-theoretic process analysis"@en ; skos:altLabel "STPA"@en ; skos:definition "Hazard analysis that models a system as a control structure and identifies unsafe control actions and the flawed feedback or process models that cause them."@en ; mlsys:specializes ; mlsys:uses ; mlsys:detects . a skos:Concept ; skos:inScheme ; skos:prefLabel "Failure modes and effects analysis"@en ; skos:altLabel "FMEA"@en ; skos:definition "Enumerating each component's possible failure modes and tracing their effects on system behavior to rank and mitigate them."@en ; mlsys:specializes ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Functional insufficiency (SOTIF)"@en ; skos:altLabel "SOTIF"@en ; skos:altLabel "safety of the intended functionality"@en ; skos:definition "Hazardous behavior arising without any fault, from limits of a specification or of a component's performance, such as perception errors under particular triggering conditions."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Safety filter"@en ; skos:definition "A runtime component that passes a proposed action when it is certified safe and otherwise replaces it with a minimally different safe action."@en ; mlsys:implements ; mlsys:uses ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Control barrier function"@en ; skos:altLabel "CBF"@en ; skos:definition "A function whose sign defines a safe set and whose derivative condition yields input constraints that keep the system state inside that set."@en ; mlsys:enables ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Simplex architecture"@en ; skos:definition "Pairing an unverified high-performance controller with a verified safety controller and a decision module that switches to the safety controller before safety is lost."@en ; mlsys:implements ; mlsys:uses ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Minimal-risk condition"@en ; skos:altLabel "MRC"@en ; skos:altLabel "safe fallback"@en ; skos:altLabel "fallback ladder"@en ; skos:definition "A predefined reduced-capability state the system reaches when it can no longer operate safely, often via an ordered ladder of fallback behaviors."@en ; mlsys:specializes ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Guardrails"@en ; skos:altLabel "content safety filter"@en ; skos:altLabel "safety classifier"@en ; skos:definition "Runtime checks on model inputs and outputs that detect and block, rewrite, or escalate content or actions violating a stated safety policy."@en ; mlsys:detects ; mlsys:specializes ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Scenario-based testing"@en ; skos:definition "Validating a system against a structured catalog of operating scenarios and their parameter ranges, with coverage measured over the operating domain."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Simulation-based falsification"@en ; skos:definition "Searching scenario parameters, often guided by temporal-logic robustness, to find counterexamples that violate a safety specification."@en ; mlsys:uses ; mlsys:enables ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Formal verification"@en ; skos:definition "Mathematically proving that a model of a system satisfies a formal specification for all behaviors within stated assumptions."@en ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Neural network verification"@en ; skos:definition "Proving or refuting input-output properties of a trained network over a bounded input region, such as local robustness, using exact or bound-propagation methods."@en ; mlsys:specializes ; mlsys:measures ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Rare-event safety validation"@en ; skos:definition "Estimating or bounding failure probabilities too small to observe in feasible exposure, using importance sampling, extreme-value models, or zero-failure statistical bounds."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "X-in-the-loop testing"@en ; skos:altLabel "SIL"@en ; skos:altLabel "HIL"@en ; skos:altLabel "software-in-the-loop"@en ; skos:altLabel "hardware-in-the-loop"@en ; skos:definition "A qualification ladder that runs the real software, then real hardware, against simulated or recorded environments before physical trials."@en ; mlsys:measures ; mlsys:measures ; mlsys:specializes ; mlsys:implements . a skos:Concept ; skos:inScheme ; skos:prefLabel "Goal Structuring Notation"@en ; skos:altLabel "GSN"@en ; skos:definition "A graphical notation for safety cases that links goals, strategies, solutions (evidence), context, and assumptions into an explicit argument structure."@en ; mlsys:implements ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Functional safety"@en ; skos:definition "Absence of unreasonable risk from malfunctioning electrical or electronic systems, achieved through a standardized safety lifecycle such as IEC 61508 or ISO 26262."@en ; mlsys:requires ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Safety integrity level"@en ; skos:altLabel "SIL"@en ; skos:altLabel "ASIL"@en ; skos:definition "A graded risk class assigned to a safety function that sets the rigor of development, verification, and tolerable failure rate required."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Safety performance indicator"@en ; skos:altLabel "SPI"@en ; skos:definition "A field-measured metric tied to a safety-case claim whose threshold violation signals that the argument may no longer hold."@en ; mlsys:measures ; mlsys:requires ; mlsys:detects ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Group fairness metrics"@en ; skos:altLabel "demographic parity"@en ; skos:altLabel "equalized odds"@en ; skos:altLabel "predictive parity"@en ; skos:definition "Statistics comparing selection rates, error rates, or calibration of a model's decisions across protected groups."@en ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Individual fairness"@en ; skos:definition "A requirement that individuals who are similar with respect to the task receive similar outcomes under a stated similarity metric."@en ; mlsys:trades-off ; mlsys:specializes ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Fairness criteria incompatibility"@en ; skos:definition "The result that calibration and error-rate parity across groups cannot generally hold together when base rates differ, forcing an explicit choice among criteria."@en ; mlsys:limits ; mlsys:limits ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Bias mitigation"@en ; skos:definition "Interventions on data (pre-processing), training objective (in-processing), or decisions (post-processing) that reduce measured disparities, usually at some cost in other metrics."@en ; mlsys:trades-off ; mlsys:implements ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Feature attribution"@en ; skos:altLabel "SHAP"@en ; skos:altLabel "LIME"@en ; skos:altLabel "integrated gradients"@en ; skos:altLabel "saliency"@en ; skos:definition "Post-hoc explanation assigning each input feature a contribution to a specific prediction."@en ; mlsys:increases ; mlsys:detects . a skos:Concept ; skos:inScheme ; skos:prefLabel "Counterfactual explanation"@en ; skos:definition "An explanation stating the smallest feasible change to an input that would change the model's decision."@en ; mlsys:enables ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Inherently interpretable model"@en ; skos:definition "A model whose structure, such as a sparse linear model, short rule list, or shallow tree, can be directly inspected to understand its predictions."@en ; mlsys:alternative-to ; mlsys:enables ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model card"@en ; skos:definition "A standardized document reporting a model's intended use, evaluation results across conditions and groups, limitations, and ethical considerations."@en ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Datasheet for datasets"@en ; skos:altLabel "data card"@en ; skos:definition "A standardized document recording a dataset's motivation, composition, collection process, preprocessing, intended uses, and maintenance."@en ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Contestability and recourse"@en ; skos:definition "The ability of an affected person to understand, challenge, and obtain review or reversal of a consequential automated decision."@en ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Automation bias"@en ; skos:altLabel "automation complacency"@en ; skos:altLabel "ironies of automation"@en ; skos:definition "The tendency of human supervisors to over-rely on automated outputs and to lose the vigilance and skill needed to catch their errors."@en ; mlsys:reduces ; mlsys:limits ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Authority handover"@en ; skos:altLabel "takeover"@en ; skos:altLabel "shared autonomy"@en ; skos:definition "Transfer of control between an automated system and a human operator, bounded by human reaction time, situational awareness, and mode clarity."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "AI risk management"@en ; skos:altLabel "NIST AI RMF"@en ; skos:altLabel "ISO/IEC 42001"@en ; skos:definition "An organizational process to govern, map, measure, and manage risks of AI systems throughout their lifecycle."@en ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Algorithmic impact assessment"@en ; skos:definition "A structured pre-deployment analysis of who an AI system could affect, how, and with what severity, used to decide mitigations or whether to deploy."@en ; mlsys:part-of ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Algorithmic audit"@en ; skos:definition "An internal or independent examination of an AI system's development records and behavior against stated requirements, standards, or law."@en ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Risk-tiered AI regulation"@en ; skos:altLabel "EU AI Act"@en ; skos:definition "Law that classifies AI uses by risk level and imposes obligations, such as conformity assessment, documentation, and human oversight, scaled to that level."@en ; mlsys:requires ; mlsys:requires ; mlsys:requires ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Frontier safety framework"@en ; skos:altLabel "responsible scaling policy"@en ; skos:altLabel "preparedness framework"@en ; skos:definition "A developer policy that ties dangerous-capability evaluation thresholds to required safeguards before further training or deployment of frontier models."@en ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Context compaction"@en ; skos:definition "Replacing older parts of an agent's context with summaries or structured records so a long task fits the model's input budget."@en ; mlsys:reduces ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Long-context degradation"@en ; skos:altLabel "lost in the middle"@en ; skos:altLabel "context rot"@en ; skos:definition "Decline in a model's use of relevant information as context grows or as that information sits away from the context's start or end."@en ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Long-term agent memory"@en ; skos:definition "A persistent store an agent writes to and retrieves from across calls and sessions, with policies for consolidation, expiry, and invalidation."@en ; mlsys:uses ; mlsys:enables ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Function calling"@en ; skos:altLabel "tool calling"@en ; skos:altLabel "tool use"@en ; skos:definition "A model emitting a structured invocation of a declared tool, which the harness validates, executes, and returns as an observation."@en ; mlsys:requires ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Tool discovery protocol"@en ; skos:altLabel "Model Context Protocol"@en ; skos:altLabel "MCP"@en ; skos:definition "An open client-server protocol through which agent hosts discover and invoke tools, resources, and prompts exposed by external servers."@en ; mlsys:enables ; mlsys:increases ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Computer use"@en ; skos:altLabel "GUI agent"@en ; skos:altLabel "browser agent"@en ; skos:definition "An agent operating graphical software through screenshots or accessibility trees and synthesized keyboard and pointer input."@en ; mlsys:increases ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Agent harness"@en ; skos:altLabel "agent loop"@en ; skos:altLabel "agent runtime"@en ; skos:definition "The runtime that repeatedly calls a model, dispatches its tool actions, appends observations, and decides whether the trajectory continues, pauses, or ends."@en ; mlsys:produces ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Agent execution budget"@en ; skos:definition "Limits on turns, tokens, wall-clock time, or spend for a trajectory, with stopping rules applied when a limit or completion criterion is reached."@en ; mlsys:determines ; mlsys:limits ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Action approval gate"@en ; skos:altLabel "human-in-the-loop approval"@en ; skos:definition "A harness checkpoint that holds a specific proposed action until an authorized human or policy approves it, binding the approval to that exact action."@en ; mlsys:implements ; mlsys:mitigates ; mlsys:produces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Durable execution"@en ; skos:definition "Recording each workflow step's intent and result in a persistent log so execution resumes after failure without repeating completed steps."@en ; mlsys:requires ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Record and replay"@en ; skos:altLabel "deterministic replay"@en ; skos:definition "Re-executing a trajectory from its log by substituting recorded model outputs and tool results, to reproduce, debug, or test a divergence."@en ; mlsys:uses ; mlsys:enables ; mlsys:requires ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Suspend and resume"@en ; skos:definition "Persisting a trajectory while it waits on a human, timer, or external event, then resuming with a validated state and authority."@en ; mlsys:enables ; mlsys:requires ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Multi-agent orchestration"@en ; skos:altLabel "orchestrator-worker"@en ; skos:altLabel "delegation"@en ; skos:definition "Structuring a task as delegated subtasks among agents under a topology, such as orchestrator-worker or peer, with typed handoffs and result merging."@en ; mlsys:causes ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Shared agent state"@en ; skos:definition "Task state read and written by several agents, requiring concurrency control such as leases or optimistic versioning to avoid lost or conflicting updates."@en ; mlsys:uses ; mlsys:part-of ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Multi-agent coordination failure"@en ; skos:definition "Task failure arising from interaction among agents, such as mis-specified delegation, lost context at handoff, conflicting actions, or premature termination."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Correlated agent errors"@en ; skos:definition "Errors shared across agents built on the same model or prompts, which make voting, debate, or cross-checking less independent than assumed."@en ; mlsys:limits ; mlsys:limits ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Robot perception"@en ; skos:definition "Models that convert raw sensor streams into task-relevant scene estimates, such as detections, segmentation, depth, and pose, under latency limits."@en ; mlsys:increases ; mlsys:requires ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Sensor fusion"@en ; skos:definition "Combining measurements from multiple sensors or modalities, at raw, feature, or decision level, into one estimate more accurate or robust than any single source."@en ; mlsys:requires ; mlsys:requires ; mlsys:increases ; mlsys:uses ; mlsys:mitigates ; mlsys:uses ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Sensor time synchronization"@en ; skos:definition "Aligning clocks and timestamps across sensors, compute, and actuators so measurements can be associated with the physical instant they describe."@en ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Recursive state filtering"@en ; skos:altLabel "Kalman filter"@en ; skos:altLabel "EKF"@en ; skos:altLabel "particle filter"@en ; skos:definition "Updating a state estimate and its uncertainty by alternating model-based prediction with measurement correction at each step."@en ; mlsys:implements ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Simultaneous localization and mapping"@en ; skos:altLabel "SLAM"@en ; skos:definition "Jointly estimating a platform's pose and a map of an unknown environment from its own sensor observations."@en ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Learned world model"@en ; skos:definition "A learned model that predicts future observations or latent states conditioned on actions, used for planning, policy training, or simulation."@en ; mlsys:enables ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Motion planning"@en ; skos:definition "Computing a collision-free path or trajectory from a start to a goal configuration using search, sampling, or trajectory optimization."@en ; mlsys:requires ; mlsys:requires ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model predictive control"@en ; skos:altLabel "MPC (control)"@en ; skos:definition "Repeatedly solving a finite-horizon constrained optimal control problem from the current state and applying only the first input."@en ; mlsys:uses ; mlsys:requires ; mlsys:specializes ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Real-time scheduling"@en ; skos:altLabel "rate-monotonic"@en ; skos:altLabel "EDF"@en ; skos:altLabel "response-time analysis"@en ; skos:altLabel "hard real-time deadline"@en ; skos:definition "Assigning processor time to periodic and sporadic tasks so that each meets its deadline, with schedulability established by analysis."@en ; mlsys:requires ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Mixed-criticality partitioning"@en ; skos:definition "Running functions of different assurance levels on shared hardware while preventing lower-criticality work from delaying or corrupting higher-criticality functions."@en ; mlsys:enables ; mlsys:requires ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Vision-language-action model"@en ; skos:altLabel "VLA"@en ; skos:altLabel "robot foundation model"@en ; skos:definition "A policy built on a pretrained vision-language backbone that maps images and language instructions to robot actions."@en ; mlsys:increases ; mlsys:uses ; mlsys:specializes ; mlsys:uses ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Teleoperation data"@en ; skos:definition "Robot demonstrations collected by humans driving the robot remotely, whose cost, coverage, and action-label fidelity bound imitation-learned policies."@en ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Action chunking"@en ; skos:definition "Predicting a short sequence of future actions per policy inference, decoupling model invocation rate from control rate and reducing compounding error."@en ; mlsys:increases ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Domain randomization"@en ; skos:definition "Training in simulation over randomized visual and physical parameters so the policy treats the real world as one more variation."@en ; mlsys:enables ; mlsys:uses ; mlsys:specializes ; mlsys:mitigates ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Thread hierarchy"@en ; skos:altLabel "grid/block/warp"@en ; skos:definition "The decomposition of a kernel into grids of thread blocks and warps that share on-chip memory and synchronize at defined scopes."@en ; mlsys:uses ; mlsys:implements ; mlsys:uses ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Memory coalescing and bank conflicts"@en ; skos:altLabel "coalesced access"@en ; skos:definition "Access patterns that merge a warp's loads into few memory transactions, or that serialize on shared-memory banks."@en ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Occupancy and latency hiding"@en ; skos:altLabel "occupancy"@en ; skos:definition "The ratio of resident to maximum warps, and how oversubscribing threads hides memory latency at the cost of registers and shared memory per thread."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Asynchronous data movement"@en ; skos:altLabel "warp specialization"@en ; skos:altLabel "TMA"@en ; skos:definition "Hardware-assisted bulk copies into on-chip memory and producer-consumer warp roles that overlap loads with matrix-engine compute."@en ; mlsys:implements ; mlsys:uses ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Matrix-engine programming"@en ; skos:altLabel "tensor core programming"@en ; skos:altLabel "MMA"@en ; skos:definition "Issuing tile-level matrix-multiply-accumulate instructions with the layouts, fragment shapes, and precisions a tensor-core-class unit requires."@en ; mlsys:enables ; mlsys:uses ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Embedding-table sharding"@en ; skos:altLabel "DLRM sharding"@en ; skos:definition "Partitioning very large sparse embedding tables row-, column-, or table-wise across devices, usually model-parallel beside data-parallel dense layers."@en ; mlsys:enables ; mlsys:requires ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Embedding cache and tiering"@en ; skos:altLabel "embedding offload"@en ; skos:definition "Keeping frequently accessed embedding rows in accelerator memory and cold rows in host memory or SSD, exploiting the skew of lookups."@en ; mlsys:increases ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Parameter server"@en ; skos:altLabel "PS architecture"@en ; skos:definition "An architecture in which workers pull and push model parameters from sharded server processes, common for sparse recommendation models and asynchronous training."@en ; mlsys:uses ; mlsys:alternative-to ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Multi-stage retrieval and ranking"@en ; skos:altLabel "ranking funnel"@en ; skos:altLabel "candidate generation"@en ; skos:definition "Candidate generation followed by progressively costlier ranking stages under one latency budget, the dominant architecture for recommendation and search."@en ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Multimodal model"@en ; skos:altLabel "vision-language model"@en ; skos:altLabel "VLM"@en ; skos:definition "A model that consumes or produces several modalities, such as text, images, audio, or video, through modality encoders feeding a shared backbone."@en ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Modality encoder"@en ; skos:altLabel "visual tokens"@en ; skos:altLabel "vision encoder"@en ; skos:definition "A network that turns images, audio, or video into token sequences for a shared backbone; visual tokens often dominate prefill cost and KV size."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Multimodal encoder disaggregation"@en ; skos:altLabel "encode-prefill-decode disaggregation"@en ; skos:definition "Scheduling modality encoders separately from prefill and decode and caching encoder outputs so image or audio inputs do not stall token generation."@en ; mlsys:reduces ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Automatic parallelization"@en ; skos:altLabel "auto-sharding"@en ; skos:altLabel "automatic parallelism"@en ; skos:definition "Compiler-derived multi-device partitioning of a single-device program, found by propagating sharding annotations or by searching parallelism strategies under a cost model."@en ; mlsys:produces ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Sharding annotation and propagation"@en ; skos:altLabel "SPMD partitioning"@en ; skos:altLabel "GSPMD-style sharding"@en ; skos:definition "Annotating a few tensors with layouts over a device mesh and letting the compiler infer the remaining layouts and insert required collectives, yielding one SPMD program."@en ; mlsys:implements ; mlsys:produces ; mlsys:requires ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Device mesh"@en ; skos:definition "A logical multidimensional arrangement of devices whose axes are assigned to parallelism dimensions, usually aligned so the busiest axis maps to the fastest interconnect."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Composable function transformations"@en ; skos:altLabel "grad/vmap/jit"@en ; skos:altLabel "functional transformations"@en ; skos:definition "Higher-order operations that map pure array functions to new functions, such as differentiation, vectorization, compilation, and sharding, and compose in any order."@en ; mlsys:implements ; mlsys:requires ; mlsys:uses ; mlsys:implements . a skos:Concept ; skos:inScheme ; skos:prefLabel "Distributed task and actor runtime"@en ; skos:altLabel "actor model runtime"@en ; skos:altLabel "Ray-style runtime"@en ; skos:definition "A runtime exposing stateless remote tasks and stateful actors as future-returning calls, scheduled across a cluster with a shared object store and lineage-based fault tolerance."@en ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Single- vs. multi-controller execution"@en ; skos:definition "Whether one coordinating process issues work to all devices or every host runs the same program independently, trading orchestration flexibility against dispatch scalability."@en ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Analytical performance model"@en ; skos:definition "A closed-form model composing compute, memory, communication, and queueing terms to predict a workload's time, memory, and cost on a configuration without running it."@en ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Performance simulation"@en ; skos:altLabel "system simulation"@en ; skos:definition "Predicting system behavior by executing a model of compute, memory, network, or scheduling events at a chosen granularity, trading simulation speed against detail."@en ; mlsys:alternative-to ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Simulator fidelity"@en ; skos:altLabel "validation error"@en ; skos:definition "Agreement between a performance model's or simulator's predictions and measured behavior on reference workloads, reported as validation error over stated configurations."@en ; mlsys:measures ; mlsys:measures ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Post-training quantization"@en ; skos:altLabel "PTQ"@en ; skos:altLabel "calibration set"@en ; skos:definition "Quantizing a trained model without retraining, choosing scales and rounding from a small calibration set of representative inputs."@en ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Weight-only quantization"@en ; skos:altLabel "W4A16"@en ; skos:altLabel "GPTQ"@en ; skos:altLabel "AWQ"@en ; skos:definition "Storing weights at low bit width while computing in higher-precision activations, cutting memory traffic for bandwidth-bound decoding, with error-compensating or activation-aware rounding."@en ; mlsys:specializes ; mlsys:reduces ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Activation outliers"@en ; skos:altLabel "outlier features"@en ; skos:altLabel "SmoothQuant"@en ; skos:definition "Rare, large-magnitude activation channels in large transformers that inflate quantization ranges, handled by mixed-precision decomposition or by migrating scale into weights."@en ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "KV-cache quantization"@en ; skos:altLabel "KV quantization"@en ; skos:definition "Storing attention keys and values at reduced precision, often with per-channel keys and per-token values, to raise batch size or context length under fixed memory."@en ; mlsys:reduces ; mlsys:reduces ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "N:M structured sparsity"@en ; skos:altLabel "2:4 sparsity"@en ; skos:altLabel "semi-structured sparsity"@en ; skos:definition "A fine-grained pattern keeping at most N nonzero values in every group of M, such as 2:4, that matrix engines can skip for real speedup."@en ; mlsys:requires ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "KV-cache offloading and tiering"@en ; skos:altLabel "KV offload"@en ; skos:altLabel "KV tiering"@en ; skos:definition "Moving KV-cache blocks from accelerator memory to host memory or SSD and reloading them on reuse, extending cache capacity at transfer bandwidth and latency cost."@en ; mlsys:enables ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "KV-cache transfer"@en ; skos:altLabel "KV migration"@en ; skos:definition "Shipping KV state between instances or nodes, as in disaggregated prefill and decode or cross-instance prefix reuse, overlapped with compute to hide network latency."@en ; mlsys:uses ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Compound AI system"@en ; skos:altLabel "LM program"@en ; skos:altLabel "LLM pipeline"@en ; skos:definition "An application composing multiple model calls, retrievers, and tools into one program whose components and control flow are tuned and served jointly for quality and cost."@en ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "LM program optimization"@en ; skos:altLabel "prompt optimization"@en ; skos:definition "Automatically tuning the prompts, demonstrations, or weights of each model call in a declared pipeline against an end-to-end metric rather than per call."@en ; mlsys:determines ; mlsys:requires ; mlsys:implements . a skos:Concept ; skos:inScheme ; skos:prefLabel "Multi-adapter serving"@en ; skos:altLabel "multi-LoRA serving"@en ; skos:altLabel "multi-tenant LoRA"@en ; skos:definition "Serving many low-rank adapters over one shared base model by batching requests for different adapters with heterogeneous kernels and paging adapter weights on demand."@en ; mlsys:uses ; mlsys:specializes ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Denoising-step reduction"@en ; skos:altLabel "few-step sampling"@en ; skos:altLabel "step distillation"@en ; skos:definition "Lowering network evaluations per diffusion sample with faster samplers, step distillation, or consistency models, trading sample quality against latency and cost."@en ; mlsys:uses ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Cross-step feature caching"@en ; skos:altLabel "diffusion caching"@en ; skos:definition "Reusing intermediate activations that change little between adjacent denoising steps instead of recomputing them, cutting per-step cost at bounded quality loss."@en ; mlsys:increases ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Graph-based vector index"@en ; skos:altLabel "HNSW"@en ; skos:altLabel "proximity-graph index"@en ; skos:definition "An approximate nearest-neighbor index linking vectors into a proximity graph, often hierarchical, searched by greedy traversal for high recall at high memory cost."@en ; mlsys:implements ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Partitioned and quantized vector index"@en ; skos:altLabel "IVF-PQ"@en ; skos:altLabel "product quantization"@en ; skos:definition "An index that clusters vectors into partitions probed selectively and compresses them with product quantization, trading recall for memory footprint and scan cost."@en ; mlsys:implements ; mlsys:alternative-to ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Filtered vector search"@en ; skos:altLabel "predicate-constrained ANN"@en ; skos:definition "Nearest-neighbor search restricted to records satisfying metadata predicates, where naive pre- or post-filtering loses recall or cost under selective filters."@en ; mlsys:specializes ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Vector index freshness"@en ; skos:altLabel "streaming ANN updates"@en ; skos:definition "The delay and recall cost of reflecting inserts, updates, and deletes in an ANN index, managed by incremental insertion, delta indexes, tombstones, and periodic rebuilds."@en ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Numerical nondeterminism"@en ; skos:definition "Run-to-run variation in floating-point results because non-associative reductions change order with atomics, kernel selection, parallel layout, batch composition, or hardware."@en ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Batch invariance"@en ; skos:altLabel "batch-invariant kernels"@en ; skos:definition "A kernel property that each request's output is bitwise independent of batch size and co-batched requests, required for deterministic serving and matching training to inference numerics."@en ; mlsys:mitigates ; mlsys:enables ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Checkpoint resharding"@en ; skos:altLabel "universal checkpoint"@en ; skos:altLabel "distributed checkpoint"@en ; skos:definition "Loading a checkpoint saved under one parallel layout or world size into another by remapping sharded tensors, needed for elastic restarts and training-to-serving handoff."@en ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Distributed batch data processing"@en ; skos:altLabel "MapReduce"@en ; skos:altLabel "Spark-style dataflow"@en ; skos:definition "Transforming bounded datasets with dataflow engines that partition records across workers, shuffle between stages, and re-execute failed tasks, for large-scale preprocessing and feature generation."@en ; mlsys:enables ; mlsys:enables ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Distributed tracing"@en ; skos:altLabel "request tracing"@en ; skos:altLabel "spans"@en ; skos:definition "Recording request-scoped, causally linked spans across services so that one request's latency and failures can be attributed to the components it traversed."@en ; mlsys:part-of ; mlsys:enables ; mlsys:detects . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model and agent call tracing"@en ; skos:altLabel "LLM observability"@en ; skos:altLabel "GenAI tracing"@en ; skos:definition "Tracing model calls, retrievals, and tool invocations as spans annotated with prompts, token counts, latency, and cost, so quality and spend are attributable per step."@en ; mlsys:specializes ; mlsys:measures ; mlsys:enables ; mlsys:detects . a skos:Concept ; skos:inScheme ; skos:prefLabel "Evaluation harness"@en ; skos:altLabel "eval harness"@en ; skos:definition "A system running versioned benchmark suites against model checkpoints with fixed prompts, decoding settings, and scorers, sharding and caching work so results stay reproducible and comparable."@en ; mlsys:implements ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Energy measurement methodology"@en ; skos:altLabel "power measurement"@en ; skos:definition "The declared sensor sources, sampling rate, measurement boundary, and attribution rules by which a workload's energy is measured, from on-chip counters to facility meters."@en ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Idle vs. dynamic power"@en ; skos:altLabel "static vs. dynamic power"@en ; skos:definition "Separating power drawn by provisioned but idle hardware from the incremental power of active work, which decides how shared energy is attributed to workloads."@en ; mlsys:determines ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Homomorphic-encryption inference"@en ; skos:altLabel "HE inference"@en ; skos:altLabel "encrypted inference"@en ; skos:definition "Evaluating a model on encrypted inputs with homomorphic encryption so the server never sees plaintext, at large compute overhead and with polynomial approximations of nonlinearities."@en ; mlsys:alternative-to ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Secure multi-party computation inference"@en ; skos:altLabel "secure two-party inference"@en ; skos:altLabel "SMPC inference"@en ; skos:altLabel "2PC inference"@en ; skos:definition "Jointly evaluating a model over secret-shared inputs or weights so no party learns another's data, with cost dominated by communication rounds for nonlinear layers."@en ; mlsys:alternative-to ; mlsys:mitigates ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Red teaming"@en ; skos:altLabel "AI red teaming"@en ; skos:altLabel "automated red teaming"@en ; skos:definition "Structured adversarial probing of a model or system, by people or automated attacker models, to elicit harmful, insecure, or policy-violating behavior before and after release."@en ; mlsys:detects ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Dangerous-capability evaluation"@en ; skos:altLabel "frontier capability evaluation"@en ; skos:altLabel "extreme-risk evaluation"@en ; skos:definition "Evaluating whether a model has capabilities, such as cyberoffense, biological uplift, or autonomous replication, that cross thresholds requiring added safeguards."@en ; mlsys:part-of ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Compute-in-memory"@en ; skos:altLabel "in-memory computing"@en ; skos:altLabel "processing-in-memory"@en ; skos:altLabel "PIM"@en ; skos:altLabel "analog in-memory computing"@en ; skos:definition "Performing matrix-vector operations inside or beside memory arrays, often in the analog domain, so weights need not move to separate compute units."@en ; mlsys:reduces ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Emerging non-volatile memory"@en ; skos:altLabel "ReRAM"@en ; skos:altLabel "phase-change memory"@en ; skos:altLabel "MRAM"@en ; skos:altLabel "resistive memory"@en ; skos:definition "Resistive, phase-change, and magnetic memory devices that retain state without power and can store weights densely or act as analog compute elements."@en ; mlsys:enables ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Embedded ML DSL"@en ; skos:altLabel "host-embedded DSL"@en ; skos:altLabel "staged array language"@en ; skos:definition "A domain-specific language hosted inside a general language such as Python, whose operations build a program representation that a separate compiler or runtime executes."@en ; mlsys:produces ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Differentiable programming"@en ; skos:altLabel "differentiable programming language"@en ; skos:definition "Writing general programs, including control flow, recursion, and data structures, that the language or framework differentiates end to end by automatic differentiation."@en ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Persistent kernel"@en ; skos:altLabel "megakernel"@en ; skos:altLabel "mega-kernel"@en ; skos:altLabel "persistent megakernel"@en ; skos:definition "A long-running kernel that stays resident on the device and executes many operators, or a whole forward pass, from an in-kernel work queue."@en ; mlsys:reduces ; mlsys:alternative-to ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Tensor graph superoptimization"@en ; skos:altLabel "graph substitution"@en ; skos:altLabel "TASO-style superoptimization"@en ; skos:definition "Automatically generating and verifying semantics-preserving rewrites of tensor computation graphs or kernels, then searching the rewritten space under a cost model."@en ; mlsys:part-of ; mlsys:uses ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Large-model training framework"@en ; skos:altLabel "Megatron-LM"@en ; skos:altLabel "DeepSpeed"@en ; skos:altLabel "distributed training library"@en ; skos:definition "A library that composes data, tensor, pipeline, and sharded-state parallelism with memory optimizations behind one training configuration for models exceeding one device."@en ; mlsys:implements ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Cloud service models"@en ; skos:altLabel "IaaS"@en ; skos:altLabel "PaaS"@en ; skos:altLabel "SaaS"@en ; skos:altLabel "managed ML service"@en ; skos:definition "Divisions of responsibility between cloud provider and customer, from rented virtual infrastructure to managed platforms to hosted model APIs."@en ; mlsys:determines ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data warehouse and lakehouse"@en ; skos:altLabel "data warehouse"@en ; skos:altLabel "data lake"@en ; skos:altLabel "lakehouse"@en ; skos:definition "Analytical stores of governed tables queried with SQL; a lakehouse adds transactional table formats over open files in object storage."@en ; mlsys:uses ; mlsys:enables ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Synchronous training power swings"@en ; skos:altLabel "training power fluctuations"@en ; skos:altLabel "training load swings"@en ; skos:definition "Rapid cluster-wide swings in power draw as synchronized accelerators alternate between compute and communication or stall together at checkpoints and failures."@en ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "UPS and battery energy storage"@en ; skos:altLabel "UPS"@en ; skos:altLabel "uninterruptible power supply"@en ; skos:altLabel "BESS"@en ; skos:definition "Facility batteries that carry load through outages until generators start and can also buffer fast load swings or shave demand peaks."@en ; mlsys:mitigates ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Grid interconnection"@en ; skos:altLabel "grid connection"@en ; skos:altLabel "utility interconnection"@en ; skos:definition "The utility connection that fixes how much power a site can draw; its capacity and multi-year lead times increasingly decide where AI datacenters are built."@en ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "On-site power generation"@en ; skos:altLabel "behind-the-meter generation"@en ; skos:altLabel "backup generators"@en ; skos:definition "Generators, turbines, fuel cells, or renewables at the datacenter that provide backup power or supply load alongside or instead of the grid."@en ; mlsys:alternative-to ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data cleaning"@en ; skos:altLabel "data cleansing"@en ; skos:altLabel "data repair"@en ; skos:definition "Detecting and correcting erroneous, inconsistent, duplicated, or outlying records with rules, statistics, or learned repair models before the data is used for training."@en ; mlsys:uses ; mlsys:mitigates ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Class imbalance"@en ; skos:altLabel "imbalanced data"@en ; skos:altLabel "imbalanced classes"@en ; skos:definition "A skewed label distribution in which some classes are rare, so aggregate accuracy misleads and training favors the majority classes."@en ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Class rebalancing"@en ; skos:altLabel "oversampling"@en ; skos:altLabel "undersampling"@en ; skos:altLabel "class weighting"@en ; skos:altLabel "SMOTE"@en ; skos:definition "Changing per-class sampling rates, synthesizing minority examples, or weighting the loss so rare classes contribute adequately to training."@en ; mlsys:mitigates ; mlsys:trades-off . a skos:Concept ; skos:inScheme ; skos:prefLabel "Feature hashing"@en ; skos:altLabel "hashing trick"@en ; skos:altLabel "hashed features"@en ; skos:definition "Mapping high-cardinality or open-vocabulary categorical values into a fixed number of buckets with a hash function, bounding table size at the cost of collisions."@en ; mlsys:specializes ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Feature cross"@en ; skos:altLabel "crossed features"@en ; skos:altLabel "feature conjunction"@en ; skos:definition "A synthetic feature formed from the conjunction of two or more categorical features so a model, especially a linear one, can learn their interaction."@en ; mlsys:specializes ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data governance"@en ; skos:altLabel "data stewardship"@en ; skos:definition "Policies, roles, and controls that decide who may collect, access, use, and share which data, for which purposes, and under what accountability."@en ; mlsys:determines ; mlsys:requires ; mlsys:uses ; mlsys:uses ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Consent and purpose limitation"@en ; skos:altLabel "purpose limitation"@en ; skos:altLabel "data-use consent"@en ; skos:definition "Restricting collection and use of personal data to purposes the data subject agreed to or a legal basis allows, recorded so downstream uses can be checked."@en ; mlsys:part-of ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Exploratory data analysis"@en ; skos:altLabel "EDA"@en ; skos:altLabel "data profiling"@en ; skos:definition "Summarizing and plotting distributions, correlations, and anomalies in a dataset before modeling to inform cleaning, feature design, and validation rules."@en ; mlsys:detects ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model and prediction visualization"@en ; skos:altLabel "training dashboards"@en ; skos:altLabel "embedding projection"@en ; skos:altLabel "confusion-matrix plots"@en ; skos:definition "Plotting training curves, activations, embeddings, and prediction errors such as confusion matrices to inspect what a model learned and where it fails."@en ; mlsys:uses ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model selection"@en ; skos:altLabel "model development"@en ; skos:definition "Choosing among candidate model families, architectures, and hyperparameters by validated performance under the system's accuracy, cost, and latency constraints."@en ; mlsys:uses ; mlsys:causes ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Pretraining"@en ; skos:altLabel "base-model pretraining"@en ; skos:altLabel "pretraining run"@en ; skos:definition "Training a model from scratch on a large, broad corpus with a self-supervised objective to produce a general base model for later adaptation."@en ; mlsys:uses ; mlsys:enables ; mlsys:uses ; mlsys:uses ; mlsys:enables ; mlsys:produces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Causal language modeling"@en ; skos:altLabel "next-token prediction"@en ; skos:altLabel "autoregressive language modeling"@en ; skos:definition "Training a model to predict each token from the tokens before it, maximizing the likelihood of text under a left-to-right factorization."@en ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Generative adversarial network"@en ; skos:altLabel "GAN"@en ; skos:definition "A generator and a discriminator trained against each other, so the generator learns to produce samples the discriminator cannot tell apart from data."@en ; mlsys:alternative-to ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Mode collapse"@en ; skos:definition "A generative-training failure in which the model covers only a few modes of the data distribution and produces low-diversity samples."@en ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Perplexity"@en ; skos:definition "The exponentiated average negative log-likelihood per token of held-out text, measuring how well a language model predicts that text."@en ; mlsys:specializes ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Reference-match scoring"@en ; skos:altLabel "exact match"@en ; skos:altLabel "token-overlap F1"@en ; skos:altLabel "BLEU"@en ; skos:altLabel "ROUGE"@en ; skos:definition "Scoring a generated output by exact or token-overlap agreement with reference answers; cheap and reproducible but blind to valid answers phrased differently."@en ; mlsys:specializes ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Error analysis"@en ; skos:definition "Inspecting a model's mistakes, grouping them into categories, and ranking categories by frequency and fixability to decide the next data or model change."@en ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Training diagnostics"@en ; skos:altLabel "overfit-one-batch test"@en ; skos:altLabel "learning-curve diagnosis"@en ; skos:definition "Checks such as overfitting a tiny batch, reading learning curves, and tracking gradient and activation statistics that localize bugs in training code, data, and configuration."@en ; mlsys:detects ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Request correlation keys"@en ; skos:altLabel "keyed predictions"@en ; skos:altLabel "pass-through keys"@en ; skos:definition "Passing a client-supplied key through inference unchanged so outputs of batch or asynchronous requests can be matched to their inputs regardless of order."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Stateless model serving"@en ; skos:altLabel "stateless serving function"@en ; skos:definition "Exporting a model as a stateless function from request inputs to prediction, so any replica can serve any request and replicas scale horizontally."@en ; mlsys:enables ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model gateway"@en ; skos:altLabel "LLM gateway"@en ; skos:altLabel "AI gateway"@en ; skos:definition "A proxy in front of one or more model providers that centralizes authentication, routing, rate limits, fallbacks, and usage logging for model calls."@en ; mlsys:uses ; mlsys:enables ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Semantic cache"@en ; skos:altLabel "semantic caching"@en ; skos:definition "Returning a stored model response when a new request is close in embedding space to an earlier one, trading answer fidelity for latency and cost."@en ; mlsys:uses ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Glue code"@en ; skos:definition "Supporting code written to move data into and out of general-purpose ML packages, often far larger than the model code and binding the system to package specifics."@en ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Pipeline jungle"@en ; skos:definition "Data preparation grown incrementally into a tangle of scrapes, joins, and intermediate outputs that is hard to test, trace, and change."@en ; mlsys:specializes ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Build-versus-buy decision"@en ; skos:altLabel "build vs. buy"@en ; skos:altLabel "MLOps tool selection"@en ; skos:definition "Choosing whether to build an ML platform component in house, adopt open source, or buy a managed service, weighing fit, cost, lock-in, and team capacity."@en ; mlsys:determines ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "User feedback collection"@en ; skos:altLabel "explicit feedback"@en ; skos:altLabel "implicit feedback"@en ; skos:definition "Capturing explicit signals such as ratings and corrections and implicit signals such as clicks, edits, and abandonment, for use as labels or quality signals."@en ; mlsys:uses ; mlsys:enables ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Human-AI interaction design"@en ; skos:altLabel "human-centered AI"@en ; skos:altLabel "human-AI guidelines"@en ; skos:definition "Designing how a system conveys its capabilities, uncertainty, and errors to users and lets them correct, override, or dismiss its outputs."@en ; mlsys:mitigates ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Datacenter demand response"@en ; skos:altLabel "flexible datacenter load"@en ; skos:altLabel "grid-interactive datacenter"@en ; skos:definition "Shifting, shedding, or capping datacenter power in response to grid signals, made practical by deferrable training and batch work."@en ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Prompt engineering"@en ; skos:altLabel "prompt design"@en ; skos:definition "Designing instructions, examples, and input layout for a model call to elicit desired behavior without changing the model's weights."@en ; mlsys:alternative-to ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "In-context learning"@en ; skos:altLabel "few-shot prompting"@en ; skos:altLabel "zero-shot prompting"@en ; skos:definition "A pretrained model performing a task from instructions or examples supplied in its input at inference time, with no weight updates."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Streaming speech recognition"@en ; skos:altLabel "streaming ASR"@en ; skos:altLabel "on-device speech recognition"@en ; skos:definition "Serving speech recognition incrementally: consuming audio in chunks and emitting partial transcripts within a fixed lag, with encoder state carried between chunks."@en ; mlsys:part-of ; mlsys:uses ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Voice interaction pipeline"@en ; skos:altLabel "voice agent"@en ; skos:altLabel "ASR-LLM-TTS cascade"@en ; skos:altLabel "speech-to-speech model"@en ; skos:definition "Serving a spoken interaction as one latency-budgeted pipeline of streaming speech recognition, a language model, and speech synthesis, with turn detection and interruption handling between stages."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "KV-cache eviction"@en ; skos:altLabel "KV token eviction"@en ; skos:altLabel "KV-cache sparsification"@en ; skos:altLabel "heavy-hitter retention"@en ; skos:definition "Retaining only a selected subset of cached key-value entries, chosen by attention scores, recency, or learned importance, to bound KV memory during long generation."@en ; mlsys:reduces ; mlsys:alternative-to ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Dynamic sparse attention"@en ; skos:altLabel "content-adaptive sparse attention"@en ; skos:altLabel "input-dependent sparse attention"@en ; skos:definition "Attention that computes only the query-key blocks predicted important for each input at runtime, rather than a fixed pattern, to cut long-context cost."@en ; mlsys:specializes ; mlsys:alternative-to ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "CPU-GPU hybrid inference"@en ; skos:altLabel "host-assisted inference"@en ; skos:altLabel "CPU offloaded attention"@en ; skos:definition "Splitting inference so the host CPU computes some operations, such as attention or rarely active neurons, from host memory when model or KV state exceeds accelerator memory."@en ; mlsys:alternative-to ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Heterogeneous-accelerator training and serving"@en ; skos:altLabel "heterogeneous GPU clusters"@en ; skos:altLabel "mixed-generation GPU training"@en ; skos:definition "Partitioning one training or serving job across accelerators of different types or generations, sizing stages and batch shares to each device's compute, memory, and links."@en ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Graph neural network"@en ; skos:altLabel "GNN"@en ; skos:altLabel "message-passing neural network"@en ; skos:definition "Neural network that computes node, edge, or graph representations by repeatedly aggregating features from each node's neighbors along graph edges."@en ; mlsys:specializes ; mlsys:uses ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Neighbor sampling"@en ; skos:altLabel "graph sampling"@en ; skos:altLabel "mini-batch graph sampling"@en ; skos:definition "Building GNN mini-batches from sampled subsets of each node's multi-hop neighborhood, bounding the exponential neighbor growth that makes full-neighborhood computation exceed memory."@en ; mlsys:reduces ; mlsys:increases ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Dynamic re-parallelization"@en ; skos:altLabel "parameter reallocation"@en ; skos:altLabel "runtime resharding"@en ; skos:definition "Redistributing model weights among devices at runtime so successive phases, such as generation and training or prefill and decode, each run under a better parallel layout."@en ; mlsys:uses ; mlsys:implements ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Contextual activation sparsity"@en ; skos:altLabel "dynamic activation sparsity"@en ; skos:altLabel "contextual sparsity"@en ; skos:definition "Skipping neurons, heads, or channels whose activations are predicted negligible for the current input, reducing per-token compute and weight reads at inference."@en ; mlsys:specializes ; mlsys:alternative-to ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Non-prefix KV reuse"@en ; skos:altLabel "position-independent KV caching"@en ; skos:altLabel "KV cache blending"@en ; skos:definition "Reusing precomputed KV entries for text chunks appearing at arbitrary positions across requests, selectively recomputing cross-chunk attention instead of the full prefill."@en ; mlsys:alternative-to ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "ML-aware cluster scheduling"@en ; skos:altLabel "deep-learning cluster scheduling"@en ; skos:altLabel "goodput-aware scheduling"@en ; skos:definition "Allocating cluster accelerators using ML job properties, such as scaling curves, iteration time, and parallelism choice, to optimize objectives like job completion time or goodput."@en ; mlsys:uses ; mlsys:uses ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Adaptive routing"@en ; skos:altLabel "flowlet switching"@en ; skos:altLabel "packet spraying"@en ; skos:altLabel "ECMP collision avoidance"@en ; skos:definition "Choosing network paths per packet or flowlet from current congestion, avoiding the hash collisions static ECMP suffers with few large, synchronized training flows."@en ; mlsys:reduces ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Incast congestion"@en ; skos:altLabel "incast"@en ; skos:altLabel "many-to-one congestion"@en ; skos:definition "Congestion that occurs when many senders transmit to one receiver at once, overflowing its switch port buffer and causing queueing, drops, or pause propagation."@en ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Context selection"@en ; skos:altLabel "context budget allocation"@en ; skos:altLabel "tool selection"@en ; skos:altLabel "working-set selection"@en ; skos:definition "Choosing which instructions, tool descriptions, retrieved items, state, and observations enter a model's limited context at each step, within a token budget."@en ; mlsys:uses ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Context staleness"@en ; skos:altLabel "stale observations"@en ; skos:altLabel "context invalidation"@en ; skos:definition "Observations held in an agent's context or memory that no longer reflect the environment after later actions or external changes, yet still steer decisions unless invalidated."@en ; mlsys:causes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Iterative multi-hop retrieval"@en ; skos:altLabel "agentic retrieval"@en ; skos:altLabel "multi-hop RAG"@en ; skos:definition "Retrieval that issues successive queries, each formulated from previously retrieved evidence or reasoning steps, to answer questions no single lookup resolves."@en ; mlsys:specializes ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Capability-based authorization"@en ; skos:altLabel "capability tokens"@en ; skos:altLabel "object capabilities"@en ; skos:altLabel "attenuated credentials"@en ; skos:definition "Authorization in which holding an unforgeable, scoped reference grants a specific right that can be delegated with attenuated scope and later revoked."@en ; mlsys:specializes ; mlsys:implements . a skos:Concept ; skos:inScheme ; skos:prefLabel "Ephemeral sandbox"@en ; skos:altLabel "copy-on-write workspace"@en ; skos:altLabel "disposable execution environment"@en ; skos:altLabel "microVM sandbox"@en ; skos:definition "An isolated execution environment created per task from a snapshot, often copy-on-write, and destroyed or reset afterward so no state carries across tasks or tenants."@en ; mlsys:implements ; mlsys:mitigates ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Network egress control"@en ; skos:altLabel "egress filtering"@en ; skos:altLabel "outbound allowlist"@en ; skos:definition "Restricting which external network destinations a workload or sandbox may contact, typically through default-deny allowlists, to block exfiltration and unintended remote actions."@en ; mlsys:implements ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Information-flow control"@en ; skos:altLabel "taint tracking"@en ; skos:altLabel "data-flow labels"@en ; skos:definition "Labeling data with confidentiality and integrity tags and propagating them through computation so policy can block untrusted inputs from steering privileged actions or secrets from leaking."@en ; mlsys:mitigates ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Compounding error in sequential decisions"@en ; skos:altLabel "covariate shift in imitation"@en ; skos:altLabel "exposure bias"@en ; skos:altLabel "horizon error compounding"@en ; skos:definition "Per-step errors of a learned sequential policy or generator push later inputs away from the training distribution, so failure probability grows with horizon length."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Interactive imitation learning"@en ; skos:altLabel "DAgger"@en ; skos:altLabel "dataset aggregation"@en ; skos:altLabel "recovery demonstrations"@en ; skos:definition "Imitation learning that queries an expert for corrective labels on states the learner's own policy visits and aggregates them into training data over rounds."@en ; mlsys:specializes ; mlsys:mitigates ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Loss masking and token normalization"@en ; skos:altLabel "prompt masking"@en ; skos:altLabel "assistant-only loss"@en ; skos:definition "Excluding prompt, tool-output, and other non-target tokens from the training loss, and choosing how remaining token losses are averaged across turns and sequences."@en ; mlsys:part-of ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Credit assignment"@en ; skos:altLabel "temporal credit assignment"@en ; skos:altLabel "multi-turn credit assignment"@en ; skos:definition "Attributing a delayed or episode-level reward to the individual actions, turns, or tokens that produced it, so updates reinforce the responsible decisions."@en ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Actuator limits"@en ; skos:altLabel "torque saturation"@en ; skos:altLabel "thermal duty limits"@en ; skos:definition "Bounds on the force, torque, speed, and sustained duty an actuator can deliver, set by drive current limits, gearing, and motor heating."@en ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Watchdog timer"@en ; skos:altLabel "heartbeat timeout"@en ; skos:altLabel "windowed watchdog"@en ; skos:altLabel "authority-expiry timer"@en ; skos:definition "A timer that a monitored component must refresh within a deadline or window; expiry triggers a defined safe response such as reset or revoking control authority."@en ; mlsys:part-of ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Closed-loop policy evaluation"@en ; skos:altLabel "closed-loop testing"@en ; skos:altLabel "interactive policy evaluation"@en ; skos:definition "Evaluating a policy by letting its own actions determine subsequent observations, in simulation or deployment, rather than scoring its predictions against logged data."@en ; mlsys:detects . a skos:Concept ; skos:inScheme ; skos:prefLabel "Event data recorder"@en ; skos:altLabel "black-box recorder"@en ; skos:altLabel "safety event log"@en ; skos:altLabel "authority record"@en ; skos:definition "Tamper-evident onboard logging of inputs, decisions, control authority, and safety interventions around events, retained for incident reconstruction and assurance evidence."@en ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Decision accountability record"@en ; skos:altLabel "decision ownership"@en ; skos:altLabel "accountability assignment"@en ; skos:altLabel "approval record"@en ; skos:definition "A record of who authorized each consequential decision, policy, or automated action and on what evidence, so responsibility can be assigned and reviewed later."@en ; mlsys:enables ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Agent interruption and cancellation"@en ; skos:altLabel "pause and cancel"@en ; skos:altLabel "operator interrupt"@en ; skos:altLabel "agent kill switch"@en ; skos:definition "Operator controls that pause, redirect, cancel, or terminate a running agent at defined boundaries while settling in-flight tool calls into a known state."@en ; mlsys:part-of ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Agent loop detection"@en ; skos:altLabel "semantic watchdog"@en ; skos:altLabel "livelock detection"@en ; skos:altLabel "repetition detection"@en ; skos:definition "Monitoring an agent trajectory for repeated actions, cycling states, or absent progress, and halting or escalating when the agent appears stuck."@en ; mlsys:part-of ; mlsys:detects ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Task and motion planning"@en ; skos:altLabel "TAMP"@en ; skos:definition "Planning that jointly searches discrete symbolic action sequences and continuous motions, accepting a task plan only when each step admits a feasible motion."@en ; mlsys:uses ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Multi-object tracking"@en ; skos:altLabel "MOT"@en ; skos:altLabel "data association"@en ; skos:altLabel "track management"@en ; skos:definition "Maintaining identities and states of multiple moving objects over time by associating new detections with existing tracks and starting or ending tracks."@en ; mlsys:specializes ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Safety change-impact analysis"@en ; skos:altLabel "impact analysis"@en ; skos:altLabel "safety-case maintenance"@en ; skos:definition "Determining which hazards, safety-case claims, and evidence a change to a model, software, data, or operating domain invalidates, and what must be re-verified."@en ; mlsys:requires ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Sparse tensor kernels"@en ; skos:altLabel "SpMM"@en ; skos:altLabel "SDDMM"@en ; skos:altLabel "sparse convolution"@en ; skos:definition "Kernels such as sparse-dense matrix multiply, sampled dense-dense multiply, and sparse convolution that compute over compressed nonzero structure, trading irregular access for skipped work."@en ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Local SGD"@en ; skos:altLabel "periodic model averaging"@en ; skos:altLabel "DiLoCo-style training"@en ; skos:altLabel "low-communication training"@en ; skos:definition "Data-parallel training in which workers take several local optimizer steps between parameter synchronizations, cutting communication frequency for low-bandwidth or geo-distributed settings."@en ; mlsys:specializes ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Expert replication and placement"@en ; skos:altLabel "expert load balancing"@en ; skos:altLabel "redundant experts"@en ; skos:definition "Assigning mixture-of-experts experts to devices, and replicating frequently routed ones, to balance per-device token load under skewed routing."@en ; mlsys:part-of ; mlsys:mitigates ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Memory-efficient optimizer"@en ; skos:altLabel "low-rank optimizer state"@en ; skos:altLabel "8-bit optimizer"@en ; skos:definition "Optimizers that shrink per-parameter state through factorization, low-rank projection, or low-precision storage, cutting the memory adaptive methods such as Adam otherwise require."@en ; mlsys:specializes ; mlsys:reduces ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Operator execution trace"@en ; skos:altLabel "execution graph trace"@en ; skos:altLabel "Chakra trace"@en ; skos:definition "A recorded graph of a training or inference step's compute and communication operators with their dependencies and timings, replayable for simulation and benchmarking."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Offline reinforcement learning"@en ; skos:altLabel "batch reinforcement learning"@en ; skos:altLabel "offline RL"@en ; skos:definition "Learning a policy from a fixed dataset of previously logged interactions without further environment access, limited by how well the data covers states the policy would visit."@en ; mlsys:specializes ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Byzantine-robust aggregation"@en ; skos:altLabel "robust gradient aggregation"@en ; skos:altLabel "Krum"@en ; skos:definition "Combining worker or client updates with robust statistics, such as medians, trimmed means, or Krum, so a bounded fraction of arbitrary malicious updates cannot steer training."@en ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Cluster manager"@en ; skos:altLabel "cluster orchestrator"@en ; skos:altLabel "cluster control plane"@en ; skos:altLabel "Kubernetes"@en ; skos:definition "The control plane that admits jobs, places their tasks on machines, enforces quotas and priorities, and reconciles running state toward declared intent across a cluster."@en ; mlsys:uses ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Rebound effect"@en ; skos:altLabel "Jevons paradox"@en ; skos:definition "Efficiency gains lower the cost of a service enough that total demand, and with it total energy or resource use, grows, partly or fully offsetting the savings."@en ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data cascade"@en ; skos:altLabel "compounding data defects"@en ; skos:definition "Compounding downstream harm from upstream data problems, such as poor collection or labeling, that surface late, often after deployment, and are costly to trace."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data valuation"@en ; skos:altLabel "Data Shapley"@en ; skos:altLabel "data attribution"@en ; skos:definition "Estimating each training example's or data source's contribution to model quality, for example through influence or Shapley-style attribution, to guide selection, pricing, or removal."@en ; mlsys:enables ; mlsys:measures . a skos:Concept ; skos:inScheme ; skos:prefLabel "Diffusion policy"@en ; skos:altLabel "multimodal action head"@en ; skos:altLabel "flow-matching policy"@en ; skos:definition "A robot policy that generates actions by iterative denoising, representing multimodal action distributions that regression policies average into invalid in-between actions."@en ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Sensor spoofing"@en ; skos:altLabel "transduction attack"@en ; skos:altLabel "physical sensor attack"@en ; skos:altLabel "LiDAR spoofing"@en ; skos:definition "Attacks that inject false signals into a sensor's physical channel, such as fabricated LiDAR returns or GPS and acoustic spoofing, so perception receives counterfeit measurements."@en ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Paired comparison and noninferiority testing"@en ; skos:altLabel "paired difference test"@en ; skos:altLabel "noninferiority test"@en ; skos:definition "Comparing two systems on the same items or units and testing per-item differences, including whether a candidate is no worse than a baseline by a preset margin."@en ; mlsys:reduces ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "CPU execution architecture"@en ; skos:altLabel "general-purpose CPU"@en ; skos:altLabel "host CPU"@en ; skos:altLabel "CPU vector units"@en ; skos:definition "Organization of general-purpose processor cores, SIMD vector units, and multilevel caches that runs host code, preprocessing, classical models, and CPU inference."@en ; mlsys:alternative-to ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "CPU threading and NUMA placement"@en ; skos:altLabel "thread affinity"@en ; skos:altLabel "intra-op and inter-op threads"@en ; skos:altLabel "NUMA-aware execution"@en ; skos:definition "Configuring thread pools, core affinity, and memory placement across NUMA domains so CPU operators parallelize without oversubscription or remote-memory penalties."@en ; mlsys:uses ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Dynamic shapes and shape specialization"@en ; skos:altLabel "symbolic shapes"@en ; skos:altLabel "shape bucketing"@en ; skos:altLabel "shape recompilation"@en ; skos:definition "Handling inputs whose tensor dimensions vary across calls by specializing compiled code per shape, padding into buckets, or compiling with symbolic dimensions."@en ; mlsys:part-of ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Rule-based graph rewriting"@en ; skos:altLabel "constant folding"@en ; skos:altLabel "algebraic simplification"@en ; skos:altLabel "operator substitution"@en ; skos:altLabel "graph optimization passes"@en ; skos:altLabel "graph rewriting"@en ; skos:definition "Applying hand-written semantics-preserving transformations to a computation graph, such as constant folding, dead-code elimination, operator substitution, and algebraic simplification."@en ; mlsys:part-of ; mlsys:alternative-to ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Sparse and ragged tensor representation"@en ; skos:altLabel "ragged tensor"@en ; skos:altLabel "nested tensor"@en ; skos:altLabel "jagged tensor"@en ; skos:altLabel "CSR format"@en ; skos:definition "Storing tensors with mostly zero entries or variable-length dimensions in compressed formats such as COO, CSR, or offset-indexed ragged layouts instead of padded dense arrays."@en ; mlsys:specializes ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Execution providers and operator fallback"@en ; skos:altLabel "execution provider"@en ; skos:altLabel "backend delegation"@en ; skos:altLabel "operator fallback"@en ; skos:definition "Partitioning a model graph across runtime backends by operator support, running supported subgraphs on accelerators and falling back to a CPU backend otherwise."@en ; mlsys:uses ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Large language model"@en ; skos:altLabel "LLM"@en ; skos:altLabel "foundation model"@en ; skos:altLabel "LLM workload"@en ; skos:definition "A language model used as a workload class: one whose training and serving are dominated by parameter memory, attention state that grows with context, and token-by-token decoding. Tag it only when an item is about that workload class itself, not merely because an LLM is involved."@en ; mlsys:uses ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Inference compute and memory estimate"@en ; skos:altLabel "2N FLOPs per token"@en ; skos:altLabel "KV-cache sizing"@en ; skos:definition "Approximating dense transformer inference as about two FLOPs per parameter per token, plus KV-cache bytes from layers, KV heads, head dimension, context length, batch, and precision."@en ; mlsys:measures ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Synchronous vs. asynchronous training"@en ; skos:altLabel "synchronous SGD"@en ; skos:altLabel "asynchronous SGD"@en ; skos:altLabel "async SGD"@en ; skos:definition "Choosing whether workers apply an update only after every gradient for a step is combined, or apply updates independently on possibly stale parameters."@en ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Input prefetching"@en ; skos:altLabel "data prefetching"@en ; skos:altLabel "prefetch buffer"@en ; skos:definition "Preparing upcoming batches in background workers and buffers while the accelerator processes the current batch, overlapping input delivery with computation."@en ; mlsys:part-of ; mlsys:increases ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Event-time windowing and watermarks"@en ; skos:altLabel "stream watermarks"@en ; skos:altLabel "late data handling"@en ; skos:altLabel "windowed aggregation"@en ; skos:definition "Grouping stream events into windows by when they occurred, using watermarks to estimate completeness and explicit policies for events arriving after a window closes."@en ; mlsys:part-of ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Token streaming"@en ; skos:altLabel "incremental response delivery"@en ; skos:altLabel "server-sent events"@en ; skos:definition "Delivering generated output to the client incrementally as tokens are produced, holding per-stream state and stopping generation when the client cancels or disconnects."@en ; mlsys:uses ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Serverless inference"@en ; skos:altLabel "scale-to-zero"@en ; skos:altLabel "function-as-a-service inference"@en ; skos:definition "Running models on a managed platform that provisions replicas on demand, bills per use, and can scale to zero, exposing requests to cold-start latency."@en ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "LLM request scheduling and preemption"@en ; skos:altLabel "request preemption"@en ; skos:altLabel "iteration-level scheduling"@en ; skos:definition "Ordering, prioritizing, and preempting in-flight generation requests at iteration granularity, swapping or recomputing their KV state, to meet latency targets and fairness goals."@en ; mlsys:uses ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Containerization"@en ; skos:altLabel "containers"@en ; skos:altLabel "Docker"@en ; skos:definition "Packaging a model, its code, and its dependencies as an immutable image that runs unchanged across machines. Scheduling those containers across a cluster belongs to the cluster manager."@en ; mlsys:implements ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Distributed job hang diagnosis"@en ; skos:altLabel "collective timeout"@en ; skos:altLabel "NCCL watchdog"@en ; skos:altLabel "node health checks"@en ; skos:definition "Detecting and localizing stalled distributed jobs through collective timeouts, heartbeats, per-rank progress traces, and node health checks, so faulty ranks can be isolated."@en ; mlsys:detects ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Benchmark representativeness"@en ; skos:altLabel "microbenchmark vs. end-to-end benchmark"@en ; skos:altLabel "benchmark realism"@en ; skos:definition "The degree to which a benchmark's workload, scale, and measurement boundary predict production behavior, from isolated microbenchmarks to end-to-end application runs."@en ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data and model licensing"@en ; skos:altLabel "dataset license"@en ; skos:altLabel "model license"@en ; skos:altLabel "open-weight license"@en ; skos:definition "Legal terms governing how datasets and model weights may be used, modified, redistributed, or deployed, which constrain acquisition, fine-tuning, and release decisions."@en ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Content provenance"@en ; skos:altLabel "C2PA"@en ; skos:altLabel "content credentials"@en ; skos:definition "Cryptographically signed metadata attached to media that records its origin and edit history, including model generation, so recipients can verify source claims."@en ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Fleet data engine"@en ; skos:altLabel "data engine"@en ; skos:altLabel "triggered data collection"@en ; skos:altLabel "edge-case mining"@en ; skos:altLabel "data flywheel"@en ; skos:definition "Collecting training data from deployed devices through on-device triggers that flag rare or uncertain events, uploading selected samples under bandwidth budgets for labeling and retraining."@en ; mlsys:uses ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Robotics middleware and dataflow"@en ; skos:altLabel "ROS 2"@en ; skos:altLabel "publish-subscribe middleware"@en ; skos:altLabel "DDS"@en ; skos:definition "Software connecting sensing, perception, planning, and control processes as nodes that exchange typed messages over publish-subscribe topics with configurable delivery and timing policies."@en ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Parallel physics simulation"@en ; skos:altLabel "GPU-parallel simulation"@en ; skos:altLabel "simulated environments"@en ; skos:altLabel "vectorized environments"@en ; skos:definition "Running many simulated robot environments concurrently, often batched on accelerators, so policy learning and data generation gather experience faster than physical collection."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Agent credentials and delegated authority"@en ; skos:altLabel "agent identity"@en ; skos:altLabel "delegated tokens"@en ; skos:altLabel "secret brokering"@en ; skos:definition "Issuing tool-using agents scoped, short-lived credentials that act on a user's behalf, brokering secrets outside model context so authority stays bounded and revocable."@en ; mlsys:implements ; mlsys:uses ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Graph partitioning for GNN training"@en ; skos:altLabel "graph partitioning"@en ; skos:altLabel "METIS partitioning"@en ; skos:altLabel "distributed GNN training"@en ; skos:definition "Splitting a large graph across workers to balance computation while minimizing cut edges, whose cross-partition neighbor features must be fetched remotely."@en . a skos:Concept ; skos:inScheme ; skos:prefLabel "Host-side data loading"@en ; skos:altLabel "data loader"@en ; skos:altLabel "DataLoader workers"@en ; skos:altLabel "pinned memory"@en ; skos:definition "Reading, decoding, augmenting, and collating examples on host CPU worker processes and staging batches in pinned memory so that a single machine's accelerators are not left waiting for input."@en ; mlsys:part-of ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Streaming inference"@en ; skos:altLabel "incremental inference"@en ; skos:altLabel "online streaming inference"@en ; skos:definition "Serving a model over an open stream of input, producing outputs incrementally from partial input while carrying state between chunks, rather than answering one complete request at a time."@en ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Inference API and payload serialization"@en ; skos:altLabel "prediction API"@en ; skos:altLabel "gRPC inference"@en ; skos:altLabel "tensor serialization"@en ; skos:definition "The request and response protocol of a model server, including the transport (HTTP, gRPC, streaming connections), the tensor and payload encoding, and the cost of serializing inputs and outputs."@en ; mlsys:part-of ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Request throughput"@en ; skos:altLabel "queries per second"@en ; skos:altLabel "QPS"@en ; skos:altLabel "tokens per second"@en ; skos:definition "The number of requests, or tokens, a serving system completes per unit time at a stated load and latency target."@en ; mlsys:part-of ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Latency percentiles"@en ; skos:altLabel "p50"@en ; skos:altLabel "p99"@en ; skos:altLabel "percentile latency"@en ; skos:definition "Summarizing a latency distribution by percentiles such as p50, p95, and p99, so that a service is judged by the experience of most requests and of the slow tail rather than by the mean."@en ; mlsys:measures ; mlsys:implements . a skos:Concept ; skos:inScheme ; skos:prefLabel "Cache-aware request routing"@en ; skos:altLabel "prefix-affinity routing"@en ; skos:altLabel "KV-aware routing"@en ; skos:altLabel "session affinity"@en ; skos:definition "Sending each request to the replica that already holds a reusable prefix or KV cache for it, trading load balance against cache reuse."@en ; mlsys:uses ; mlsys:trades-off . a skos:Concept ; skos:inScheme ; skos:prefLabel "Distributed job launch and process groups"@en ; skos:altLabel "rendezvous"@en ; skos:altLabel "ranks"@en ; skos:altLabel "communicator initialization"@en ; skos:altLabel "torchrun"@en ; skos:definition "Starting the processes of a distributed job, having them discover each other through a rendezvous, assigning ranks, and forming the process groups and communicators that collectives run over."@en ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Agentic workload"@en ; skos:altLabel "agent workload"@en ; skos:altLabel "multi-turn inference load"@en ; skos:definition "The load an agent places on the serving stack, including context that grows each turn, idle gaps while tools run, many short dependent model calls, and reuse of KV state across steps."@en ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Performance regression testing"@en ; skos:altLabel "continuous benchmarking"@en ; skos:altLabel "perf regression gate"@en ; skos:definition "Benchmarking every change to code, kernels, models, or configuration continuously and failing the change when throughput, latency, or memory degrades beyond a set tolerance."@en ; mlsys:uses ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Performance measurement variability"@en ; skos:altLabel "benchmark noise"@en ; skos:altLabel "warm-up effects"@en ; skos:altLabel "timing variance"@en ; skos:definition "Run-to-run variation in measured time and throughput from warm-up, asynchronous execution, clock and thermal state, and contention, handled by synchronizing, repeating runs, and reporting spread."@en ; mlsys:limits ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "PII detection and de-identification"@en ; skos:altLabel "PII scrubbing"@en ; skos:altLabel "data anonymization"@en ; skos:altLabel "redaction"@en ; skos:definition "Finding personal identifiers in training corpora, logs, and prompts and removing, masking, or pseudonymizing them before the data is stored, trained on, or shared."@en ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Gradient clipping"@en ; skos:altLabel "gradient norm clipping"@en ; skos:definition "Rescaling or truncating gradients whose norm exceeds a threshold so that a single large update cannot destabilize training."@en ; mlsys:mitigates ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Weight initialization"@en ; skos:altLabel "Xavier initialization"@en ; skos:altLabel "He initialization"@en ; skos:definition "Choosing the scale and distribution of initial parameters so that activations and gradients keep stable variance through depth at the start of training."@en ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Processing delivery semantics"@en ; skos:altLabel "exactly-once processing"@en ; skos:altLabel "at-least-once delivery"@en ; skos:altLabel "duplicate events"@en ; skos:definition "The guarantee a data pipeline gives about how often each record takes effect (at most once, at least once, or effectively once) under retries, replays, and failures, and the idempotent or transactional sinks that achieve it."@en ; mlsys:uses ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Online and offline feature consistency"@en ; skos:altLabel "feature materialization"@en ; skos:altLabel "online store consistency"@en ; skos:definition "Keeping the feature values materialized in the low-latency online store consistent with the offline values used for training, as features are recomputed, backfilled, and updated."@en ; mlsys:mitigates ; mlsys:trades-off . a skos:Concept ; skos:inScheme ; skos:prefLabel "Out-of-distribution detection"@en ; skos:altLabel "OOD detection"@en ; skos:altLabel "novelty detection"@en ; skos:definition "Flagging individual inputs that differ from the training distribution so the system can abstain, defer, or fall back instead of trusting the prediction."@en ; mlsys:detects ; mlsys:enables . a skos:Concept ; skos:inScheme ; skos:prefLabel "Model and feature compatibility"@en ; skos:altLabel "interface compatibility"@en ; skos:altLabel "tokenizer compatibility"@en ; skos:definition "Ensuring that a model version, its tokenizer and feature transformations, the schemas it reads, and the consumers of its outputs remain compatible as each is versioned and released independently."@en ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Workload identity and secrets management"@en ; skos:altLabel "secrets management"@en ; skos:altLabel "workload identity"@en ; skos:altLabel "credential rotation"@en ; skos:definition "Giving training jobs, model servers, and pipelines verifiable identities and provisioning, rotating, and revoking the keys and credentials they use, without embedding secrets in code or images."@en ; mlsys:implements . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data deletion propagation"@en ; skos:altLabel "right to erasure"@en ; skos:altLabel "deletion requests"@en ; skos:definition "Carrying a deletion request through every copy and derivative of the data, including datasets, features, caches, replicas, backups, and models trained on it."@en ; mlsys:requires ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Runtime memory fragmentation"@en ; skos:altLabel "allocator fragmentation"@en ; skos:altLabel "reserved versus allocated memory"@en ; skos:definition "Free device memory split into blocks too small to satisfy new allocations, so a job runs out of memory while reserved memory exceeds what live tensors use."@en ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Perception latency budget"@en ; skos:altLabel "sense-to-act latency budget"@en ; skos:definition "The share of a robot's sense-to-act deadline allotted to sensing and perception, which bounds model size, input resolution, and frame rate on the onboard computer."@en ; mlsys:part-of ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Onboard compute partitioning"@en ; skos:altLabel "onboard workload placement"@en ; skos:altLabel "robot compute offloading"@en ; skos:definition "Assigning perception, planning, and control workloads to the CPUs, accelerators, and microcontrollers of a robot, or offloading them, so each meets its deadline and criticality level."@en ; mlsys:uses ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Health checks and readiness probes"@en ; skos:altLabel "readiness probe"@en ; skos:altLabel "liveness probe"@en ; skos:altLabel "health endpoint"@en ; skos:definition "Periodic liveness and readiness checks that let an orchestrator or load balancer stop routing traffic to, or restart, a replica that is unhealthy or still loading its model."@en ; mlsys:enables ; mlsys:mitigates . a skos:Concept ; skos:inScheme ; skos:prefLabel "Python GIL and host-language concurrency"@en ; skos:altLabel "GIL"@en ; skos:altLabel "global interpreter lock"@en ; skos:altLabel "multiprocessing workers"@en ; skos:definition "The CPython global interpreter lock serializes Python bytecode across threads, so ML host code achieves concurrency through native kernels that release the lock, multiple processes, or asynchronous I/O."@en ; mlsys:increases ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Flash write endurance"@en ; skos:altLabel "flash wear-out"@en ; skos:altLabel "wear leveling"@en ; skos:altLabel "program/erase cycles"@en ; skos:definition "Flash memory cells wear out after a limited number of program and erase cycles, which bounds how often embedded and edge devices can write logs, checkpoints, or model updates; wear leveling spreads writes to extend life."@en ; mlsys:limits ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Energy-harvesting intermittent computing"@en ; skos:altLabel "energy harvesting"@en ; skos:altLabel "batteryless inference"@en ; skos:altLabel "intermittent inference"@en ; skos:definition "Running inference on devices powered by harvested energy that lose power unpredictably, which requires checkpointing progress to nonvolatile memory so computation resumes rather than restarts."@en ; mlsys:specializes ; mlsys:requires . a skos:Concept ; skos:inScheme ; skos:prefLabel "Memory-mapped model weights"@en ; skos:altLabel "mmap weights"@en ; skos:altLabel "shared weight mapping"@en ; skos:altLabel "zero-copy model loading"@en ; skos:definition "Mapping a weight file into virtual memory so pages load on demand and read-only weights are shared across processes, cutting load time and duplicate memory at the cost of page faults."@en ; mlsys:reduces ; mlsys:reduces ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "OS memory pressure and paging"@en ; skos:altLabel "page thrashing"@en ; skos:altLabel "low-memory killer"@en ; skos:altLabel "app memory limit"@en ; skos:altLabel "memory reclamation"@en ; skos:definition "How an operating system reclaims memory from a model process under pressure, by evicting file-backed pages, swapping or compressing memory, or killing the app, which shows up as page thrashing, latency spikes, or abrupt termination."@en ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Managed-runtime interop and garbage collection"@en ; skos:altLabel "JNI overhead"@en ; skos:altLabel "GC pauses"@en ; skos:altLabel "native interop"@en ; skos:definition "Crossing between a managed language runtime (Java, Kotlin, Swift, C#) and native inference code, where buffer copies, object pinning, and garbage-collection pauses add latency and memory."@en ; mlsys:limits ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Zero-copy sensor input via DMA"@en ; skos:altLabel "DMA input buffers"@en ; skos:altLabel "zero-copy capture"@en ; skos:altLabel "double buffering"@en ; skos:definition "Having sensors or peripherals write input directly into model input buffers by direct memory access, avoiding CPU copies between capture and inference."@en ; mlsys:alternative-to ; mlsys:reduces ; mlsys:uses . a skos:Concept ; skos:inScheme ; skos:prefLabel "Active sensor interference"@en ; skos:altLabel "sensor crosstalk"@en ; skos:altLabel "lidar interference"@en ; skos:definition "Crosstalk between active sensors, such as lidars, radars, or structured-light cameras, that emit into the same environment and corrupt each other's measurements."@en ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Network propagation delay"@en ; skos:altLabel "speed-of-light latency"@en ; skos:altLabel "round-trip time"@en ; skos:altLabel "RTT"@en ; skos:definition "The minimum latency set by the distance a signal travels at the speed of light in fiber, which bounds round trips between regions regardless of bandwidth."@en ; mlsys:limits ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Consistency-availability trade-off"@en ; skos:altLabel "CAP theorem"@en ; skos:altLabel "partition tolerance"@en ; skos:definition "Under a network partition, a replicated system must choose between answering with possibly stale state and refusing requests until replicas agree."@en ; mlsys:limits ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Decentralized gossip averaging"@en ; skos:altLabel "gossip SGD"@en ; skos:altLabel "decentralized SGD"@en ; skos:definition "Training in which each worker averages its model with a few neighbors instead of a central server or global all-reduce, so updates diffuse through the communication graph."@en ; mlsys:alternative-to ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Interrupt latency"@en ; skos:altLabel "ISR latency"@en ; skos:altLabel "interrupt service routine budget"@en ; skos:definition "The delay between a hardware event and its interrupt handler running, plus the handler's own run time, which must fit inside a real-time sensing or control budget alongside inference."@en ; mlsys:part-of ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Encoder-only, decoder-only, and encoder-decoder Transformers"@en ; skos:altLabel "encoder-only model"@en ; skos:altLabel "decoder-only model"@en ; skos:altLabel "encoder-decoder model"@en ; skos:altLabel "seq2seq Transformer"@en ; skos:definition "The three arrangements of Transformer stacks, which differ in whether inputs are processed bidirectionally, outputs are generated autoregressively, or both, and therefore in how they are trained, cached, and served."@en ; mlsys:specializes ; mlsys:determines . a skos:Concept ; skos:inScheme ; skos:prefLabel "Data transfer cost"@en ; skos:altLabel "egress cost"@en ; skos:altLabel "egress fees"@en ; skos:altLabel "network transfer cost"@en ; skos:definition "The money charged for moving data across network, zone, region, or provider boundaries, such as cloud egress fees, which can dominate the cost of distributing data, checkpoints, or responses."@en ; mlsys:increases ; mlsys:part-of . a skos:Concept ; skos:inScheme ; skos:prefLabel "Duty cycling and sleep states"@en ; skos:altLabel "duty cycle"@en ; skos:altLabel "sleep modes"@en ; skos:altLabel "active and sleep power states"@en ; skos:altLabel "low-power modes"@en ; skos:definition "Keeping a device in a low-power sleep state most of the time and waking it briefly to sense and infer, so average power is set by the active fraction and the cost of each wake-up."@en ; mlsys:uses ; mlsys:reduces . a skos:Concept ; skos:inScheme ; skos:prefLabel "Priority inversion"@en ; skos:altLabel "priority inheritance"@en ; skos:altLabel "priority ceiling"@en ; skos:definition "A high-priority task waiting on a resource held by a lower-priority task, which a medium-priority task then preempts, so the high-priority task misses its deadline unless priority inheritance or ceilings are used."@en ; mlsys:limits ; mlsys:increases . a skos:Concept ; skos:inScheme ; skos:prefLabel "Certified robustness by randomized smoothing"@en ; skos:altLabel "randomized smoothing"@en ; skos:altLabel "certified defense"@en ; skos:definition "Classifying many noise-perturbed copies of an input and taking the majority vote, which yields a provable radius within which no adversarial perturbation can change the prediction, at the cost of many inferences per input."@en ; mlsys:mitigates ; mlsys:alternative-to . a skos:Concept ; skos:inScheme ; skos:prefLabel "Sensor interconnect bandwidth"@en ; skos:altLabel "MIPI CSI-2"@en ; skos:altLabel "I2S"@en ; skos:altLabel "SPI bandwidth"@en ; skos:altLabel "camera link bandwidth"@en ; skos:definition "The data rate of the links that carry sensor data into an embedded processor (MIPI CSI-2 for cameras, I2S for audio, SPI and I2C for low-rate sensors), which bounds the resolution, frame rate, and number of sensors a device can ingest."@en ; mlsys:limits . a skos:Concept ; skos:inScheme ; skos:prefLabel "Finite buffers and overflow"@en ; skos:altLabel "buffer overflow"@en ; skos:altLabel "bounded queue"@en ; skos:altLabel "finite-capacity queue"@en ; skos:altLabel "backlog"@en ; skos:definition "A queue with bounded capacity drops or blocks arrivals once full, so its size trades memory and added delay against the loss rate under bursty input."@en ; mlsys:specializes . a skos:Concept ; skos:inScheme ; skos:prefLabel "Asynchronous on-device inference"@en ; skos:altLabel "UI-thread blocking"@en ; skos:altLabel "background inference thread"@en ; skos:altLabel "main-thread jank"@en ; skos:definition "Running inference off the application's main thread, so a model call does not block rendering and input handling, with results delivered back by callback or message."@en ; mlsys:uses ; mlsys:reduces .