AI Glossary

The percentage of predictions a model gets correct out of all predictions made. Accuracy = (Correct predictions) / (Total predictions). Simple and intuitive, but misleading when classes are imbalanced.

A mathematical function applied to the output of a neuron to introduce non-linearity. Without activation functions, a neural network would behave like a single linear equation no matter how many layers it has. Common choices: ReLU, sigmoid, tanh, softmax.

A technique that lets a model focus on the most relevant parts of its input when producing each output. Attention computes a weighted sum of "values" based on the similarity between a "query" and a set of "keys." It is the core innovation behind the Transformer architecture.

A hypothetical AI system that can perform any intellectual task a human can, including learning new skills from scratch, applying reasoning across domains, and understanding context without additional training. AGI does not currently exist.

The algorithm that trains neural networks by computing how much each weight contributed to the error, then adjusting all weights in the direction that reduces that error. It works by applying the chain rule of calculus from the output layer backwards to the input layer.

In machine learning, "bias" has two meanings. (1) Statistical bias: the error from wrong assumptions in the model, causing it to miss relevant relationships in the data (underfitting). (2) AI fairness bias: systematic unfairness baked into a model because of skewed training data or labels.

The number of training examples the model sees before updating its weights. Larger batches give more stable gradient estimates but require more memory. Smaller batches are noisier but often generalise better and train faster per epoch.

A type of supervised learning where the model predicts which category (class) an input belongs to. The output is a discrete label. Binary classification has two classes; multi-class has three or more.

An unsupervised learning technique that groups data points so that items within a group are more similar to each other than to items in other groups. No labels are required. Common algorithms include K-Means and DBSCAN.

A neural network architecture designed for grid-structured data like images. It uses convolutional filters (small sliding windows) that learn to detect local patterns like edges, textures, and shapes. Pooling layers then downsample to reduce computation.

A technique for evaluating model performance by splitting data into K folds, training K times (each time using a different fold as the validation set), and averaging the results. Gives a more reliable estimate than a single train/test split.

A subset of Machine Learning that uses neural networks with many layers (hence "deep") to learn hierarchical representations of data. Each layer transforms the data into a more abstract representation. Deep learning excels at unstructured data like images, audio, and text.

A machine learning model that makes predictions by learning a flowchart of yes/no questions about the input features. Easy to interpret and explain. Prone to overfitting on their own, but powerful when combined into ensembles (like Random Forests).

A regularisation technique where a random fraction of neurons are set to zero during each training step. This prevents neurons from co-adapting too closely, reducing overfitting and making the network more robust.

A dense vector (list of numbers) that represents a word, sentence, image, or other object in a continuous space. Items that are semantically similar end up with vectors that are close together. Embeddings allow mathematical operations like finding meaning through subtraction and addition.

One complete pass through the entire training dataset. A model typically trains for multiple epochs so it can refine its weights repeatedly. Too few epochs leads to underfitting; too many can cause overfitting.

The harmonic mean of precision and recall. F1 = 2 * (Precision * Recall) / (Precision + Recall). A single metric that balances both: it is high only when both precision and recall are high. Useful for imbalanced datasets where accuracy is misleading.

The process of using domain knowledge to create, transform, or select input variables (features) that help a machine learning model learn better. Deep learning reduces but does not eliminate the need for feature engineering.

Continuing to train a pretrained model on a new, smaller dataset so it adapts to a specific task. The model starts with knowledge from its original training and refines it rather than learning from scratch. Usually requires far less data and compute than training from scratch.

The optimisation algorithm used to train most machine learning models. It repeatedly adjusts the model's parameters in the direction that reduces the loss function, using the gradient (slope) of the loss to know which direction is "downhill."

A model's ability to perform well on new, unseen data it was not trained on. Good generalisation means the model learned underlying patterns, not just memorised training examples. The gap between training performance and test performance measures generalisation.

A configuration value set before training that controls the learning process itself, rather than being learned from data. Examples: learning rate, batch size, number of layers, number of epochs, regularisation strength. Choosing good hyperparameters is often done by trial-and-error or grid search.

An unsupervised clustering algorithm that partitions data into K clusters by iteratively assigning each point to its nearest cluster centre (centroid) and then moving each centroid to the average position of its assigned points. Sensitive to the initial centroid positions and the number K.

A very large neural network (billions of parameters) trained on vast amounts of text data to predict and generate language. LLMs learn grammar, facts, reasoning patterns, and style from the statistics of language. Examples: GPT-4, Claude, Gemini, LLaMA.

A hyperparameter that controls how large a step the model takes in the direction of the gradient during each weight update. Too large and the model overshoots the minimum and diverges. Too small and training takes extremely long or gets stuck.

A mathematical function that measures how wrong the model's predictions are. Training aims to minimise the loss. Different tasks use different loss functions: Mean Squared Error for regression, Cross-Entropy for classification, and many others.

A subset of AI where systems learn from data to improve performance on a task, without being explicitly programmed with rules for that task. The three main types are supervised learning, unsupervised learning, and reinforcement learning.

In machine learning, a model is a mathematical function with learned parameters that maps inputs to outputs. The "model" is what is left after training is complete and it is ready to make predictions on new data.

An AI system designed and trained to do one specific task or a small set of related tasks. All current commercially deployed AI systems are narrow AI. They cannot transfer knowledge to unrelated domains without retraining.

A computational model loosely inspired by the brain, consisting of layers of interconnected nodes (neurons). Each connection has a weight. The network learns by adjusting these weights so its outputs match the training labels. The foundation of deep learning.

Scaling numerical features so they fall within a similar range (e.g. 0 to 1, or with mean 0 and standard deviation 1). Prevents features with large values from dominating the learning process and helps gradient descent converge faster.

When a model learns the training data too well, including its noise and random fluctuations, and therefore performs poorly on new data. An overfit model has memorised rather than generalised. Solutions include more data, regularisation, dropout, and early stopping.

The internal values of a model that are learned during training -- primarily the weights and biases of the neurons. "Number of parameters" is commonly used to describe the size of a model. GPT-4 is estimated to have over 1 trillion parameters.

Of all the times the model predicted "positive," how many were actually positive? Precision = True Positives / (True Positives + False Positives). High precision means when the model says yes, it is usually right. Optimise for precision when false alarms are costly.

The practice of crafting inputs (prompts) to language models in ways that consistently produce better, more accurate, or more useful outputs. Techniques include few-shot examples, chain-of-thought instructions, role assignment, and output format constraints.

A technique that improves LLM answers by first retrieving relevant documents from a knowledge base, then giving the retrieved text to the model as context before generating a response. Reduces hallucination and allows up-to-date knowledge without retraining.

Of all the actual positive cases, how many did the model correctly identify? Recall = True Positives / (True Positives + False Negatives). High recall means the model misses very few real positive cases. Optimise for recall when missing a positive case is costly.

A type of supervised learning where the model predicts a continuous numerical value rather than a category. The output is a number on a scale. Common algorithms include Linear Regression and Ridge Regression.

Techniques that reduce overfitting by penalising model complexity. L1 regularisation (Lasso) pushes many weights to exactly zero. L2 regularisation (Ridge) shrinks all weights towards zero. Dropout is a regularisation technique specific to neural networks.

A type of machine learning where an agent learns to make decisions by taking actions in an environment and receiving rewards or penalties. The agent learns a policy that maximises cumulative reward over time. Used in game-playing AI, robotics, and recommendation systems.

A type of machine learning where the model learns from labelled training examples: each input comes paired with the correct output. The model learns to map inputs to outputs by minimising the difference between its predictions and the true labels.

An activation function used in the output layer of multi-class classifiers that converts raw scores (logits) into a probability distribution that sums to 1. The class with the highest probability is the model's prediction.

The basic unit of text that a language model processes. Tokens are not always whole words: they can be syllables, punctuation, or word fragments. GPT models split text into subword tokens. "Unbelievable" might be split into ["Un", "believ", "able"].

Reusing a model trained on one task as the starting point for a different but related task. The pretrained model has already learned useful general features (edges, textures, language patterns) that do not need to be relearned from scratch on the new task.

A neural network architecture introduced in 2017 ("Attention Is All You Need") built entirely on attention mechanisms. It processes all tokens in parallel (unlike RNNs) and uses self-attention to build context-aware representations. The foundation of GPT, BERT, and all modern LLMs.

The labelled (or unlabelled) examples the model learns from during training. The quality, size, and diversity of training data is often the single biggest factor in model performance. Garbage in, garbage out.

A held-out portion of data (separate from training and test sets) used to tune hyperparameters and monitor for overfitting during training. It gives an honest signal of how the model is learning without contaminating the final test evaluation.

An ordered list of numbers representing a point or direction in multi-dimensional space. In machine learning, features are represented as vectors; model weights are vectors; embeddings are vectors. Most of deep learning is matrix and vector arithmetic.

A numerical value on a connection between neurons that determines how strongly one neuron influences the next. During training, weights are continuously adjusted to reduce prediction error. The "knowledge" of a neural network is encoded in its weights.