AI Glossary

Definitions of the AI terms you'll encounter most often at BYU-Idaho — from alignment and tokens to context windows and RAG.

This is a glossary of terms commonly used in Artificial Intelligence.

• Artificial Intelligence (AI): A broad field of computer science focused on building systems that perform tasks typically requiring human intelligence, such as perception, reasoning, learning, planning, and language understanding.
• Machine Learning (ML): A subset of AI where models learn patterns from data to make predictions or decisions, improving performance on a task without being explicitly programmed with fixed rules.
• Deep Learning: A subset of ML that uses multi-layer neural networks to learn hierarchical representations, often excelling in vision, speech, and natural language tasks.
• Neural Network: A computational model composed of interconnected "neurons" (parameters organized in layers) that transforms inputs into outputs through learned weights, enabling pattern recognition and function approximation.
• Large Language Model (LLM): A neural network trained on large-scale text (and sometimes code) to predict and generate language, enabling capabilities like summarization, question answering, translation, and reasoning-like behaviors.
• Foundation Model: A large, broadly trained model that can be adapted to many downstream tasks via prompting, fine-tuning, or tool use, serving as a general-purpose base for multiple applications.
• Frontier Model: A state-of-the-art foundation model at the leading edge of capability, typically large, expensive to train, and evaluated against advanced benchmarks for reasoning, coding, multimodal understanding, and safety.
• Generative AI: AI systems that create new content such as text, images, audio, video, or code by learning the underlying structure of training data and sampling plausible outputs.
• Prompt: The input instructions and context provided to a model to steer its output, including system constraints, user requests, examples, and retrieved or tool-produced context.
• Prompt Engineering: The practice of designing prompts (structure, instructions, examples, constraints) to reliably elicit desired behavior from a model without changing model weights.
• Context Window: The maximum amount of input (tokens) a model can attend to at once, which bounds how much conversation history, documents, and instructions can be included in a single inference.
• Token: A unit of text used by models (often subwords or character groups) that determines how input and output length are measured and billed, and affects context window usage.
• Embedding: A dense numerical vector representation of content (text, images, etc.) that captures semantic meaning, enabling similarity search, clustering, and retrieval.
• Vector Database (Vector Store): A storage system optimized for indexing and searching embeddings, supporting fast approximate nearest neighbor queries used in semantic search and retrieval pipelines.
• Retrieval-Augmented Generation (RAG): An architecture where a model retrieves relevant external information (e.g., from a vector store or search index) and uses it as grounded context to generate more accurate, up-to-date, or source-based outputs.
• Fine-Tuning: A training approach that updates a pre-trained model's weights on task-specific data to improve performance, consistency, or domain adaptation beyond what prompting alone achieves.
• Inference: The process of running a trained model to produce outputs from inputs (e.g., generating text from a prompt), typically involving decoding strategies like greedy, beam, or sampling methods.
• Agent: A system that uses a model to plan and execute multi-step work toward a goal, often with memory, tools (APIs), and control logic to observe results, adapt, and iterate.
• Tool Use (Function Calling): A capability where a model selects and invokes external functions or APIs (search, database queries, code execution) to obtain information or take actions, then incorporates results into its response.
• Orchestration: The coordination layer that manages multi-step AI workflows, including routing, tool selection, retries, state/memory management, safety checks, and integration with other services to deliver reliable outcomes.
• Reinforcement Learning (RL): A learning paradigm where an agent learns to choose actions by interacting with an environment to maximize cumulative reward over time.
• Reinforcement Learning from Human Feedback (RLHF): A technique that uses human preference judgments to train a reward model and then optimize a model's behavior toward outputs humans rate as better aligned or higher quality.
• Supervised Learning: A machine learning approach where a model learns a mapping from inputs to labeled outputs using example pairs (x, y), aiming to generalize to new inputs.
• Unsupervised Learning: A machine learning approach where a model learns structure from unlabeled data, such as clusters, latent factors, or compressed representations.
• Self-Supervised Learning: A training approach where labels are derived from the data itself (e.g., predicting masked or next tokens), enabling large-scale learning without manual annotation.
• Transfer Learning: Reusing knowledge learned from one task or dataset to improve performance on another related task, often by starting from a pre-trained model.
• Multimodal Model: A model that can process and/or generate multiple data types (such as text, images, audio, video), enabling cross-modal understanding and generation.
• Computer Vision: The AI subfield focused on enabling machines to interpret and reason about images and video, including tasks like detection, segmentation, and recognition.
• Natural Language Processing (NLP): The AI subfield focused on analyzing, understanding, and generating human language, including tasks like parsing, classification, translation, and summarization.
• Speech Recognition (ASR): Automatic Speech Recognition: converting spoken audio into written text, typically using acoustic and language modeling techniques.
• Text-to-Speech (TTS): A system that converts text into spoken audio, often aiming for natural prosody, clarity, and speaker style control.
• Attention Mechanism: A neural network component that dynamically weights the importance of different input parts when producing an output, enabling models to focus on the most relevant context.
• Transformer: A neural network architecture built primarily around attention mechanisms, widely used in modern language and multimodal models for its scalability and performance.
• Training: The process of optimizing model parameters using data and an objective function (loss), typically via gradient-based methods, to improve task performance.
• Loss Function: A mathematical measure of error between model outputs and targets, used to guide training by indicating how parameters should be updated.
• Gradient Descent: An optimization method that iteratively updates parameters in the direction that reduces the loss, often using variants like SGD, Adam, or AdamW.
• Overfitting: A failure mode where a model learns patterns specific to training data (including noise) and performs poorly on new, unseen data.
• Generalization: A model's ability to perform well on unseen inputs by learning robust patterns rather than memorizing training examples.
• Hallucination: A behavior where a generative model produces plausible-sounding but incorrect or unsupported content, often due to missing context, weak grounding, or ambiguous prompts.
• Alignment: The set of methods and objectives used to steer model behavior toward human goals, safety constraints, and acceptable norms, reducing harmful or undesired outputs.