AI Glossary

Accuracy

The fraction of predictions that are correct. In classification, high accuracy can hide poor results on rare classes.

Reference: scikit-learn: Metrics and scoring

Activation function

A function applied to a layer's outputs. Nonlinear activations let stacked layers represent relationships that a purely linear network cannot.

Reference: PyTorch: Build the Neural Network

Agent

A system that uses a model to choose actions, use tools and work toward a goal over multiple steps.

Reference: Anthropic: Building effective agents

Alignment

Training and system design intended to make a model's behaviour match desired goals, instructions or preferences.

Reference: Hugging Face: Reinforcement learning from human feedback

API

An interface through which software uses another system. An AI API can accept prompts and return generated outputs.

Reference: MDN: API

Artificial general intelligence (AGI)

AI with broad competence across many cognitive tasks, often compared with human performance. Definitions and proposed thresholds differ.

Reference: Levels of AGI

Artificial intelligence (AI)

The field of building systems that perform tasks such as learning, perception and reasoning. Machine learning is one approach.

Reference: Google Cloud: What is artificial intelligence?

Attention

An operation that combines information from different positions using learned relevance scores. Self-attention relates positions within the same sequence.

Reference: Attention Is All You Need

Autoregressive generation

Generating a sequence one step at a time, with each new token conditioned on the preceding ones.

Reference: Hugging Face: Text generation

Backpropagation

Computing gradients backward through a network so an optimizer can adjust its parameters.

Reference: PyTorch: Optimizing Model Parameters

Batch

A group of examples processed together. Batch size affects training throughput, memory use and the gradient estimate.

Reference: PyTorch: Optimizing Model Parameters

Beam search

A decoding method that keeps several candidate sequences at each step and ranks their extensions by a scoring rule.

Reference: Hugging Face: Generation strategies

Benchmark

A standardized set of tasks and scoring rules for comparing models. Its score describes performance on those tasks.

Reference: Hugging Face: Evaluate

Bias

Bias can mean a systematic skew in model outputs, or an additive learned offset in a network layer.

Reference: Google: ML Fundamentals

Chain of thought (CoT)

Intermediate reasoning steps generated before an answer. Producing such steps can improve performance on tasks requiring several deductions.

Reference: Chain-of-Thought Prompting

Chatbot

An application with a conversational interface. It may use a language model; whether it acts autonomously is a separate question.

Reference: Anthropic: Building effective agents

Checkpoint

Saved model state. A training checkpoint can also include optimizer state and progress, allowing training to resume.

Reference: TensorFlow: Training checkpoints

Classification

Predicting a category or set of categories, such as whether a message is spam. Contrast with predicting a numerical value in regression.

Reference: scikit-learn: Metrics and scoring

Context window

The maximum sequence a model can process in one request, usually measured in tokens. Input and generated tokens compete for that space.

Reference: Google: Generative AI Glossary

Cross-entropy

A loss that penalizes low predicted probability for the target outcome. Next-token training commonly minimizes cross-entropy over the correct tokens.

Reference: PyTorch: CrossEntropyLoss

Dataset

A collection of examples used for training or evaluation. Separate splits serve different purposes, such as learning and validation.

Reference: Hugging Face: Dataset features

Decoder

A network component that produces outputs from representations. An autoregressive Transformer decoder predicts new tokens using the preceding context.

Reference: Hugging Face: Transformer Architectures

Decoding

The rule for selecting output tokens from a model's predicted probabilities, such as choosing the most likely token or sampling.

Reference: Hugging Face: Text generation

Deep learning

Machine learning using neural networks with multiple layers, allowing successive layers to build increasingly complex representations.

Reference: PyTorch: Build the Neural Network

Diffusion model

A generative model trained to reverse a noising process. Generation typically refines noise into structured content over several steps.

Reference: Hugging Face: Diffusion pipelines

Distillation

Training a student model using a teacher model's outputs or internal signals, often to transfer capabilities into a smaller model.

Reference: Hugging Face: Knowledge distillation

Dropout

Randomly setting some activations to zero during training to reduce overfitting. Ordinary inference uses the network without this random dropping.

Reference: Keras: Dropout

Embedding

A learned vector representation of an item, such as text. Retrieval systems compare embeddings to find related content.

Reference: Retrieval-Augmented Generation

Encoder

A network component that turns input into representations. A text Transformer encoder typically lets each position attend to both earlier and later positions.

Reference: Hugging Face: Transformer Architectures

Epoch

One pass through a training dataset. Streaming or repeated datasets may instead use a fixed number of steps per epoch.

Reference: Keras: Training and evaluation

Evaluation

Measuring a model's behaviour on chosen tasks using metrics, human assessment or both. The test setup determines what the result means.

Reference: Hugging Face: Evaluate

Feed-forward network (FFN)

A network whose computations flow forward without recurrent connections. Transformer blocks typically apply an FFN independently at each token position.

Reference: PyTorch: Build the Neural Network

Few-shot prompting

Giving a model a few examples of the desired task in its prompt, without updating its weights.

Reference: Hugging Face: Prompt engineering

Fine-tuning

Further training of an existing model on selected data to adapt its behaviour or capabilities.

Reference: Hugging Face: How Transformers work

Foundation model

A model trained on broad data that can serve as a starting point for many downstream tasks or applications.

Reference: On the Opportunities and Risks of Foundation Models

Generative AI

AI that produces content, such as text, images or audio, from patterns learned during training.

Reference: Google: Generative AI Glossary

Gradient

A vector of derivatives showing how a scalar quantity, such as loss, changes with each parameter.

Reference: PyTorch: Automatic Differentiation

Gradient descent

An optimization method that updates parameters in the direction opposite to the loss gradient, scaled by a learning rate.

Reference: PyTorch: Automatic Differentiation

Graphics processing unit (GPU)

A processor designed for highly parallel computation. Neural networks use GPUs for operations such as matrix multiplication during training and inference.

Reference: NVIDIA: CUDA Programming Guide

Grounding

Connecting generated answers to external evidence, such as retrieved documents or search results, rather than relying solely on learned weights.

Reference: Google Cloud: Grounding overview

Hallucination

Generated content that is factually false or unsupported by the supplied information, even when presented fluently and confidently.

Reference: Google: Generative AI Glossary

Hyperparameter

A configuration choice, such as layer count or learning rate, set outside the ordinary parameter-learning process.

Reference: Google: ML Fundamentals

In-context learning

Adapting to a task using instructions or examples in the current context, without changing the model's weights.

Reference: Hugging Face: Prompt engineering

Inference

Running a trained model to produce predictions or outputs. Ordinary inference does not update the model's learned parameters.

Reference: Hugging Face: How Transformers work

Instruction tuning

Fine-tuning on instructions paired with desired responses, teaching a model to respond to requests in useful ways.

Reference: Hugging Face: Supervised Fine-Tuning

KV cache

Stored key and value tensors from attention. Reusing them avoids recalculating those tensors for earlier tokens during generation. See why memory matters.

Reference: Hugging Face: How caching works

Large language model (LLM)

A language model with a large number of parameters, trained on extensive data. There is no universal size threshold.

Reference: Hugging Face: How Transformers work

Latent space

A space of learned representations of data. A latent diffusion model generates in this space before decoding the result into an image.

Reference: Hugging Face: Diffusion pipelines

Learning rate

A setting controlling the scale of parameter updates during training. It can change according to a schedule.

Reference: PyTorch: Automatic Differentiation

Logits

Raw model scores before conversion to probabilities. Language models produce a score for each candidate next token.

Reference: PyTorch: Softmax

LoRA

Low-rank adaptation: fine-tuning small added matrices while keeping the original weights fixed, reducing the number of trainable parameters.

Reference: Hugging Face: LoRA

Loss

A numerical objective used to train a model. Lower loss means better performance on that objective, not necessarily every useful task.

Reference: PyTorch: Optimizing Model Parameters

Machine learning (ML)

Methods that learn patterns or decision rules from data rather than requiring every rule to be written explicitly.

Reference: Google: Machine Learning Glossary

Mixture of experts (MoE)

An architecture that routes inputs to selected expert networks. Total parameter count can greatly exceed the parameters used for each token.

Reference: Hugging Face: Mixture of Experts Explained

Model

A computational structure with learned parameters that maps inputs to predictions or outputs. A chatbot or agent builds an application around it.

Reference: Google: Machine Learning Glossary

Model Context Protocol (MCP)

A standard protocol connecting AI applications to external tools, data sources and reusable prompts through servers.

Reference: Model Context Protocol: Introduction

Multimodal model

A model that handles more than one kind of data, such as text and images. Supported input and output types can differ.

Reference: Google: Generative AI Glossary

Neural network

A model built from connected layers of numerical operations, whose learned parameters transform inputs into useful representations and outputs.

Reference: Attention Is All You Need

Normalization

Rescaling values to control their numerical range or statistics. Layer normalization operates on each example's activations rather than across a batch.

Reference: Layer Normalization

Open weights

Model weights made available for download and local use under a stated licence. This alone does not make the whole system open source.

Reference: Open Source Initiative: Open weights

Optimizer

An algorithm, such as Adam, that updates model parameters using gradients and its own update rules.

Reference: PyTorch: Optimizing Model Parameters

Overfitting

Learning details of the training data that do not generalize, producing a model that performs worse on unseen examples.

Reference: Google: Machine Learning Glossary

Parameters and weights

The numerical values learned during training. Weights are parameters; the parameter count measures model size, not intelligence.

Reference: PyTorch: Optimizing Model Parameters

Perplexity

Exponentiated average negative log-probability of the observed tokens. Lower values mean better prediction on the same evaluation data and tokenization.

Reference: Hugging Face: Perplexity

Positional encoding

Information representing token positions or relative distances, allowing attention to distinguish different sequence orders.

Reference: Attention Is All You Need, §3.5

Precision

Among examples predicted positive, the fraction that are actually positive. It measures how often a positive prediction is correct.

Reference: scikit-learn: Metrics and scoring

Prefill

Processing the input context before generating new tokens. In a cached Transformer, this also builds the input's KV cache.

Reference: Hugging Face: How caching works

Pretraining

The initial broad training phase that gives a model its general capabilities before more targeted adaptation.

Reference: Hugging Face: How Transformers work

Prompt

Input supplied to guide a model, including instructions, examples or task data. Applications may assemble it from several sources.

Reference: Anthropic: Building effective agents

Prompt caching

Reusing previously computed state for a matching input prefix across requests, saving repeated prefill work.

Reference: Anthropic: Prompt caching

Prompt injection

An attempt to redirect an AI system through adversarial instructions, including instructions hidden in documents or tool results it processes.

Reference: Anthropic: Prompt injections

Quantization

Representing weights or activations with fewer bits to reduce memory use and potentially computation, with possible changes in accuracy. Try our browser model.

Reference: Hugging Face: Quantization

Reasoning model

A model trained to solve tasks through intermediate reasoning, often spending extra generation steps before its final answer.

Reference: DeepSeek-R1: Reasoning via reinforcement learning

Recall

Among examples that are actually positive, the fraction the model identifies. High recall means fewer missed positives.

Reference: scikit-learn: Metrics and scoring

Regression

Predicting a numerical value, such as temperature or a house price, rather than choosing a category.

Reference: scikit-learn: Metrics and scoring

Regularization

Techniques that discourage overfitting, for example by penalizing large weights or applying dropout during training.

Reference: Keras: Regularizers

Reinforcement learning (RL)

Learning a policy for choosing actions to maximize expected reward. Language-model training can use rewards assigned to generated responses.

Reference: Google: Machine Learning Glossary

Reinforcement learning from human feedback (RLHF)

Using human feedback to define a reward signal for reinforcement learning, often by training a reward model on preference comparisons.

Reference: Hugging Face: Reinforcement learning from human feedback

Retrieval-augmented generation (RAG)

Retrieving relevant information and supplying it to a model for generation. Retrieved documents need not be incorporated into the model's weights.

Reference: Retrieval-Augmented Generation

Self-supervised learning

Learning from targets derived from the data itself, such as predicting the next token in text, without manually labelling each example.

Reference: Google: Self-supervised learning

Small language model (SLM)

A language model small enough for a chosen resource budget. Like LLM, the name has no fixed parameter cutoff. Build one with us.

Reference: Hugging Face: How Transformers work

Softmax

A function that converts scores into positive values summing to one, commonly used as probabilities over candidate tokens.

Reference: PyTorch: Softmax

Speculative decoding

Accelerating generation by proposing several tokens with a cheaper draft process and verifying them with the target model.

Reference: Hugging Face: Generation strategies

Supervised learning

Learning from examples paired with target outputs, such as images with labels or prompts with desired responses.

Reference: Google: ML Fundamentals

Synthetic data

Artificially generated examples used for training or evaluation, for instance responses produced by another model.

Reference: Hugging Face: Synthetic Data Generator

Temperature

A setting that reshapes token probabilities before sampling. Lower values concentrate choices; higher values flatten the distribution.

Reference: Hugging Face: Text generation

Tensor

A numerical array with zero or more dimensions. Neural networks represent inputs, weights and intermediate results as tensors.

Reference: PyTorch: Tensors

Token

A unit a model processes, such as a word fragment, character or byte. A token is not necessarily a word.

Reference: Hugging Face: Text generation

Tokenizer

The component that converts text into token IDs and decodes token sequences back into text.

Reference: Hugging Face: Text generation

Tool calling

A model emits a structured request for a function or service. The surrounding application executes it and returns the result.

Reference: Anthropic: Handle tool calls

Top-k sampling

Sampling the next token only from the k highest-scoring candidates, after renormalizing their probabilities.

Reference: Hugging Face: Generation parameters

Top-p sampling

Sampling from the smallest set of highest-probability tokens whose cumulative probability reaches p. Also called nucleus sampling.

Reference: Hugging Face: Generation parameters

Training

Adjusting a model's parameters using data and an optimization objective. This is when the model learns, rather than simply runs.

Reference: PyTorch: Optimizing Model Parameters

Training data

Examples used to update a model's parameters. Performance on these examples alone does not show how well the model generalizes.

Reference: Keras: Training and evaluation

Transfer learning

Reusing knowledge learned on one task or dataset for another, often by adapting a pretrained model instead of training from scratch.

Reference: Keras: Transfer learning and fine-tuning

Transformer

A neural-network architecture built around attention and feed-forward layers. Many language models generate text with a decoder-only Transformer.

Reference: Attention Is All You Need

Unsupervised learning

Learning structure from data without supplied target labels, for example grouping similar examples into clusters.

Reference: scikit-learn: Unsupervised learning

Validation data

Held-out examples used to assess progress and choose settings during development, rather than directly update model parameters.

Reference: Keras: Training and evaluation

Vector database

A database that stores vectors and supports similarity search, often used to retrieve text through its embeddings.

Reference: Google Cloud: Vector Search

Vision-language model (VLM)

A model that combines visual and language information, for example to answer questions about an image.

Reference: Hugging Face: Image-text-to-text

Zero-shot prompting

Asking a model to perform a task without providing demonstrations of that task in the prompt.

Reference: Hugging Face: Prompt engineering