AI & Machine Learning Glossary

An information criterion that supports model selection by balancing model fit against model complexity.

A method used to test whether there are meaningful differences among the means of multiple groups.

A classical model for time series forecasting that combines autoregressive and moving-average components.

A generative summarization approach that rewrites source text to produce more natural and dense summaries.

A boosting method that turns weak learners into a strong ensemble by focusing increasingly on hard examples.

A popular optimization algorithm that combines adaptive learning rates with momentum-like behavior.

A parameter-efficient approach that inserts small modules into the base model to enable task adaptation.

An early attention approach that compares query and context representations through a learnable combination function.

A clustering algorithm that forms clusters through message passing around representative exemplars.

A table structure that stores summarized results derived from detailed data in order to accelerate analytical queries.

A feature structure that summarizes lower-level records into higher-level signals meaningful for modeling.

A task focused on recognizing action classes from human or object motion in video.

A task that attempts to predict a future action from a partially observed video stream before it fully unfolds.

An approach that aims to optimize labeling cost by selecting the most useful or uncertain examples for annotation.

A data-efficiency approach in which the model selects the most informative examples and requests labels from a human or expert source.

A task focused on locating and labeling specific events within an audio stream over time.

A task focused on predicting what environment or context an audio recording comes from.

The hypothesis that argues there is a meaningful effect, difference, or relationship in the data, in contrast to the null hypothesis.

A design approach that facilitates object detection by using predefined candidate boxes of different scales and aspect ratios.

The process of transforming personal data so that it can no longer be linked back to a specific individual.

The process of adding meaningful labels, notes, or markings to data so that it becomes usable for model training.

An approach that predicts sentiment for specific aspects or entity dimensions rather than overall sentiment.

A mechanism that enables a model to learn which parts of the input deserve more focus during prediction.

A control mechanism that determines which positions a model may or may not attend to during attention computation.

A matrix structure that numerically represents how much each element in a sequence attends to the others.

A neural architecture that learns low-dimensional representations by compressing and reconstructing data.

A deep learning approach that learns normal patterns and detects anomalies through reconstruction error.

An approach that learns lower-dimensional representations of data through neural-network-based compression.

A quasi-Newton optimization method that approximates second-order information to achieve efficient convergence.

A model selection criterion that applies a stronger penalty for complexity while evaluating model fit.

A classical sequence-labeling scheme that marks entity boundaries with beginning, inside, and outside tags.

A tree-based method suited to incremental and memory-efficient clustering on large datasets.

A data augmentation strategy based on translating target-language data back into the source language to create synthetic parallel data.

The process of reprocessing historical periods or filling missing historical data gaps after the fact.

An ensemble approach that improves stability by training multiple models on bootstrap samples and combining their outputs.

An approach that improves trust by explicitly showing the source passages supporting the generated answer.

An approach that adapts a model to the terminology and usage style of specific domains such as law, healthcare, or finance.

Representation structures adapted to the terminology of specific domains such as law, healthcare, or finance rather than general language.

A fine-grained face analysis task focused on estimating where a person is looking from eye direction.

An autonomous or semi-autonomous system that perceives its environment and selects actions to achieve its goals.

A quality measure indicating how consistently different annotators make similar decisions on the same data.

An operation that finds distinctive and repeatable local points in an image to support matching and alignment tasks.

The task of automatically identifying the key terms and phrases that best represent a text.

A task that matches field names with their corresponding values in a document to create structured data.

An approach that incorporates external language model knowledge to make speech recognition output more linguistically accurate.

An optimization step-size selection approach that determines how far to move along a chosen direction.

The task of recognizing entity spans such as people, organizations, locations, and dates within text.

An observation or value that deviates noticeably from the general pattern of the dataset.

An approach that extracts subject-relation-object structures from text without relying on a predefined relation schema.

An approach that aims for a model not only to detect known objects but also to handle unknown ones as a separate category.

A detection approach that can perceive a broader object world using natural language labels instead of fixed class lists.

An approach that enables a model to flag unseen classes as unknown instead of assigning them an overconfident incorrect label.

An open-standard table structure that supports versioning, transactions, and metadata management for large-scale data lake tables.

The condition in which a model has a parameter capacity far larger than the amount of available data.

A situation where a model learns the training data too closely and performs poorly on new data.

An approach that increases the number of minority-class examples to make them more visible in the dataset.

The core capability of transforming raw inputs into meaningful structures so the system can interpret its environment.

The core statistical distinction between the full target group and the subset selected from it for analysis.

The process of deriving new conclusions from available knowledge, rules, or observations and grounding decisions.

A core stream-health metric that expresses the delay gap between produced events and consumed events.

The operation of joining multiple continuous data streams by key and time logic to create meaningful event context.

An approach that groups continuous data streams into defined time or event intervals for computation.

A mechanism that determines when speech has truly ended in order to provide correct response timing in streaming ASR systems.

A real-time speech synthesis approach that begins generating audio with low latency without waiting for the full text.

An approach for ingesting continuously generated data in real time or near real time.

A processing approach based on handling continuously arriving data events with low latency.

An approach that splits rare words into smaller meaningful pieces to balance vocabulary size and coverage.

The conceptual space that contains all possible states, solution paths, and action combinations of a problem.

An architectural connection that allows information to bypass certain layers and improves training stability.

An approach in which the model uses tools such as computation, search, or external system calls to produce more accurate results.

An approach that reduces the number of majority-class examples to produce a more balanced class distribution.

The primary partitioning field used to split warehouse tables into logical segments.