CAIP (AIP-210) key terms

A training approach where a model learns from labeled input-output pairs, adjusting its parameters to minimize the gap between predicted and true answers so it can generalize to new data.

A training approach that finds structure in unlabeled data, such as clusters, density regions, or lower-dimensional representations, without any target answers being supplied during learning.

A paradigm where an agent learns a policy by taking actions in an environment and using scalar rewards as feedback, balancing exploration of new actions against exploitation of known good ones.

The known target value or category attached to a training example, which a supervised model is trained to predict for unseen inputs at inference time.

An individual measurable input variable used by a model to make predictions, such as a customer's age, a pixel value, or a token count in a document.

The practice of creating, transforming, and selecting input variables so that a model can learn patterns more effectively, for example via scaling, encoding categories, or deriving ratios.

A failure mode where a model memorizes training-set noise and performs well on seen data but poorly on new data, typically signaled by a widening gap between training and validation error.

A condition where a model is too simple or insufficiently trained to capture the underlying pattern, producing high error on both training data and unseen data.

The tension between a model's tendency to miss real patterns (bias) and its sensitivity to random fluctuations in the training sample (variance), which guides model complexity choices.

A set of techniques that penalize model complexity, such as L1 or L2 weight penalties and dropout, to reduce overfitting and improve generalization to unseen data.

A resampling procedure that splits data into multiple folds, training and validating repeatedly so performance estimates are more stable and less dependent on one particular split.

Partitioning a dataset into separate portions for fitting parameters, tuning hyperparameters, and final unbiased evaluation, keeping the test set untouched until the very end.

A modeling error where information that would not be available at prediction time, such as target-derived features or test statistics, contaminates training and inflates measured performance.

A configuration value set before training, such as learning rate, tree depth, or number of layers, that controls how the learning algorithm behaves rather than being learned from data.

The systematic search over configuration values, using grid, random, or Bayesian methods, evaluated on validation data to select settings that maximize generalization performance.

A supervised task predicting a discrete category or probability distribution over classes, evaluated with metrics such as accuracy, precision, recall, F1, and area under the ROC curve.

A supervised task predicting a continuous numeric value, evaluated with metrics such as mean absolute error, root mean squared error, and coefficient of determination.

An unsupervised technique that groups similar records together based on a distance or similarity measure, without predefined labels, to reveal natural segments in data.

Techniques such as principal component analysis or autoencoders that compress many input variables into fewer informative dimensions, aiding visualization, speed, and noise reduction.

A table cross-tabulating predicted versus actual classes, showing true positives, false positives, false negatives, and true negatives from which many classification metrics are computed.

Precision is the share of positive predictions that are correct; recall is the share of actual positives that are found. Their balance matters when class costs differ.

The harmonic mean of precision and recall, giving a single balanced metric that is useful when classes are imbalanced and both false positives and false negatives matter.

The verified reference answer or measured reality against which model predictions and evaluation metrics are compared, ideally produced by reliable human annotation or instrumentation.

The human or automated process of assigning labels to raw data, such as tagging images or rating text, whose quality and consistency directly bound achievable model performance.

A model composed of layered units that apply weighted sums and nonlinear activations, trained by backpropagation to learn hierarchical representations from raw inputs.

The subfield of machine learning using neural networks with many layers, which learn feature hierarchies automatically and excel on perceptual data such as images, audio, and text.

The algorithm that computes gradients of a loss with respect to each network parameter by applying the chain rule backward through the computation graph, enabling gradient-based updates.

An iterative optimization method that updates parameters in the direction that reduces a loss function, controlled by a learning rate and often computed on mini-batches.

Reusing a model pretrained on a large general dataset and adapting it to a narrower target task, often by fine-tuning some layers, which reduces data and compute requirements.

Continuing training of a pretrained model on task-specific or domain-specific data, adjusting some or all weights so its behavior better matches the target use case.

A large model pretrained on broad data that can be adapted to many downstream tasks through prompting, fine-tuning, or retrieval augmentation rather than training from scratch.

A transformer-based model trained on vast text corpora to predict and generate language, capable of instruction following, summarization, and question answering when appropriately prompted.

The practice of designing instructions, examples, and context supplied to a generative model to reliably elicit accurate, well-formatted, and safe outputs for a task.

An architecture that retrieves relevant documents from an external index and inserts them into a model's prompt, grounding answers in current or proprietary knowledge and reducing fabrication.

A confident but incorrect or fabricated model output not supported by source data, mitigated through grounding, retrieval, verification steps, and human review.

A store optimized for embedding vectors that supports similarity search, commonly used to retrieve semantically related passages for recommendation, search, and retrieval-augmented generation.

A dense numeric vector representing an item's meaning, learned so that semantically similar items land near each other and can be compared with distance or cosine similarity.

The discipline of operationalizing machine learning: versioning data and models, automating training and deployment, monitoring in production, and governing the lifecycle collaboratively across teams.

Degradation of production performance as input distributions or input-output relationships change over time, detected by monitoring prediction quality and data statistics against baselines.

A catalog that tracks model versions, artifacts, metadata, and approval stages, enabling reproducible deployment, rollback, and auditability across environments.

Automated pipelines that build, test, and deploy models and their surrounding code and data transformations, including validation gates that catch regressions before release.

Continuous tracking of live model inputs, outputs, latency, and quality metrics, with alerting thresholds so that drift, bias, or failures are caught and remediated promptly.

A design pattern where people review, correct, or approve model outputs at defined checkpoints, improving safety and quality while generating feedback for retraining.

A framework of principles and practices addressing fairness, transparency, privacy, safety, and accountability so AI systems behave ethically and meet stakeholder and regulatory expectations.

The study and mitigation of systematic disadvantage in model outcomes across groups, addressed through representative data, fairness metrics, subgroup testing, and documented tradeoffs.

Methods such as feature importance, partial dependence, and local surrogate explanations that help stakeholders understand why a model produced a given prediction and trust or challenge it.

Protecting personal and sensitive information throughout collection, training, and inference using minimization, consent, de-identification, encryption, and access controls aligned to regulation.

The policies, roles, and controls governing data quality, lineage, access, retention, and compliance, ensuring AI inputs are trustworthy, auditable, and lawfully used.

A shared platform that computes and serves curated features consistently for training and online inference, reducing duplication and preventing training-serving skew.

A mismatch between how features or preprocessing are computed during training versus at inference, causing production accuracy to fall below validated results.