Artificial Intelligence

A very large neural network — usually a transformer — trained on huge amounts of text to predict the next token. The basis of modern chat assistants and AI coding tools.

Building systems that improve at a task by learning from data instead of being explicitly programmed for every case.

A family of machine learning models loosely inspired by the brain — layers of simple units that, together, can approximate complex functions.

Learning a function from labelled examples — the most widely-deployed flavour of machine learning.

The two distinct phases of a machine-learning model's life — learning its parameters (training) and using them to make predictions (inference).

The neural-network architecture, introduced in 2017, that powers modern large language models, image generation, and most of contemporary AI.

The algorithm that computes how much each weight in a neural network contributed to the error — applying the chain rule layer by layer in reverse so gradient descent knows which way to adjust every parameter.

The field of teaching computers to interpret images and video — classification, detection, segmentation, generation.

A neural-network architecture that uses convolution to detect local patterns — edges, textures, shapes — in grid-like data such as images; the workhorse of computer vision for a decade.

A learned vector representation of an item — a word, an image, a user, a product — where geometric distance roughly equals semantic similarity.

The optimisation algorithm that trains almost every neural network — iteratively nudge each parameter in the direction that reduces the loss.

The field of getting computers to understand, generate, and work with human language — from spam filters and translation to the large language models behind modern chatbots.

A branch of machine learning where an agent learns to act in an environment by trial-and-error, optimising a reward signal.

A neural-network technique that lets a model weigh which parts of its input matter most for each part of its output — the key idea behind transformers and modern language models.

A model that splits data by asking yes/no questions at each node, forming a tree of decisions that is transparent, interpretable, and the building block of powerful ensemble methods.

The original artificial neuron — a single unit that takes weighted inputs, applies a threshold, and outputs a binary decision. The simplest learnable classifier and the ancestor of modern neural networks.

The craft of designing inputs to a language model to reliably elicit the desired output — through clear instructions, examples, role-setting, and structured reasoning techniques.

A generative model that learns to remove noise from corrupted data, step by step — the approach behind Stable Diffusion, DALL-E, and Sora that produces startlingly realistic images and video.

Continuing to train a pre-trained model on a smaller, task-specific dataset — adapting general capabilities to a narrow domain at a fraction of the cost of training from scratch.

A probabilistic model for sequences where an unobserved (hidden) state evolves according to a Markov chain, and each state emits an observable symbol — the foundation of speech recognition and sequence labelling.

AI systems that process and generate across multiple modalities — text, images, audio, and video — in a single model, enabling tasks like image captioning, visual question answering, and audio transcription.

A technique that improves language model answers by first fetching relevant documents from a knowledge base, then generating a response conditioned on both the query and the retrieved content.

A classifier that finds the widest possible margin between classes — supported by the training points closest to the boundary, with kernels extending it to non-linear problems.