Browse A-Z — all 284 terms
A
Comparing two versions to see which performs better.
Assigning responsibility for an AI system's actions.
A function adding non-linearity so networks learn complex patterns.
Boosts many weak learners into one strong classifier.
An adaptive optimiser widely used to train neural networks.
Inputs crafted to fool a model into errors.
Software that acts toward goals, using tools and memory.
An independent check of an AI system's risks and compliance.
Unfair outcomes caused by skewed data or design.
Principles guiding fair and responsible AI use.
The rules and processes for overseeing AI systems.
Laws and rules governing how AI is built and used.
Research into keeping advanced AI systems harmless.
A period when AI funding and interest sharply declined.
A step-by-step procedure a computer follows to solve a problem.
Making AI goals match human values and intentions.
A way for software to request a model's services.
AI that can learn and do any intellectual task humans do.
Software that performs tasks needing human-like intelligence.
Hypothetical AI far smarter than humans in every field.
A method letting models focus on the most relevant inputs.
A network that compresses input then rebuilds it.
B
The method that adjusts weights by sending errors backwards.
Text as counts of words, ignoring order.
Training many models on data samples and averaging them.
A group of examples processed together in one training step.
Normalising layer inputs to speed up training.
A bidirectional transformer model built for understanding language.
Balancing simple models against overly complex ones.
Training models in sequence so each fixes earlier errors.
A rectangle marking an object's location in an image.
C
Releasing a model to a small group first.
A gradient boosting library that handles categorical data well.
Asking a model to reason step by step.
Software that converses with users in text or speech.
OpenAI's conversational AI built on GPT models.
A saved copy of a model during training.
When one class has far fewer examples than another.
Predicting which category an example belongs to.
Anthropic's family of conversational AI models.
AI services hosted and run on cloud platforms.
On-demand computing power and storage over the internet.
Systems that mimic human thought to solve complex problems.
AI that lets machines understand images and video.
A table showing correct and wrong predictions by class.
A package holding code and its dependencies to run anywhere.
Finding images by their visual content, not keywords.
The amount of text a model can read at once.
A layer that scans input with small filters.
A network that learns spatial patterns, mainly in images.
Who owns content an AI system generates.
A large collection of texts used for training or study.
A loss that measures how wrong predicted probabilities are.
Testing a model on different data splits to check reliability.
NVIDIA's platform for running code on GPUs.
D
OpenAI's model that creates images from text.
Labelling data so models can learn from it.
Creating extra training examples by modifying existing ones.
Fixing errors and gaps in data before use.
When live data shifts away from training data.
Rules for collecting, storing, and using data.
When training data includes information it should not have.
A collection of examples used to train or test a model.
Finds dense clusters and marks sparse points as noise.
Splits data by asking yes/no questions in a tree.
Machine learning with many-layered neural networks.
Realistic fake media made by AI.
Guessing how far each part of a scene is.
A model that creates images by reversing a noising process.
A tool that builds and runs software in containers.
Randomly switching off neurons during training to prevent overfitting.
E
Halting training when validation stops improving.
Running AI models on local devices instead of the cloud.
A numeric vector that captures the meaning of data.
Combining several models for a better prediction.
Linked qubits sharing one state instantly.
One full pass through the entire training dataset.
Software that mimics a human expert using rules and knowledge.
Making AI decisions understandable to humans.
F
A balance of precision and recall in one number.
Identifying a person from their facial features.
Treating all groups equitably in AI decisions.
A measurable input property a model uses to predict.
Creating useful inputs from raw data for a model.
Pulling useful patterns from raw images.
Choosing the most useful inputs and dropping the rest.
Learning a task from only a few examples.
Further training a pre-trained model for a specific task.
A large general model adapted to many different tasks.
A layer where every neuron links to the next.
G
A model that finds clusters as overlapping bell curves.
Google's family of multimodal AI models.
Two networks competing, one generating and one judging.
AI that creates new text, images, code, or audio.
A method that learns word vectors from global word counts.
A free cloud notebook for running Python and AI code.
A transformer model family pre-trained to generate text.
A chip that runs many calculations in parallel.
The direction and size of change that reduces loss.
Builds trees one by one, each fixing the last one's errors.
An optimisation method that reduces errors step by step.
Trying many hyperparameter combinations to find the best.
The correct answers used to judge a model.
A simpler recurrent cell with gates.
H
Confident but false output presented as if true.
Builds nested clusters by merging or splitting groups.
A platform sharing pre-trained models, datasets, and tools.
A human reviewing AI decisions before they act.
A setting chosen before training that controls learning.
I
Writing a sentence that describes an image.
Assigning a label to a whole image.
A numeric vector representing an image's visual meaning.
Cleaning and resizing images before a model reads them.
Identifying objects, people, or scenes in images.
Dividing an image into regions by object or meaning.
Learning from examples given in the prompt, without retraining.
Using a trained model to make predictions on new input.
Labelling each separate object instance in an image.
Software that perceives its environment and acts on it.
Detects outliers by isolating unusual points in trees.
J
A prompt that tricks a model past its safety rules.
An interactive tool for writing and running code.
K
Groups data into K clusters by nearest centre.
Predicts from the K closest examples in data.
A platform for data science competitions and datasets.
A simple high-level library for building neural networks.
A network of facts showing how entities relate to each other.
How knowledge is stored so machines can reason with it.
A system that manages containers at scale.
L
The correct answer attached to a training example.
A framework for building applications on top of LLMs.
A model that predicts and generates sequences of words.
A huge model trained on text to understand and generate language.
Regression that shrinks some coefficients to zero.
The delay between a request and a model's response.
A group of neurons processing data at one stage.
How big each step is when a model updates its weights.
A plan that changes the learning rate during training.
Reducing a word to its base dictionary form.
A fast gradient boosting framework built for large datasets.
Finds a line that best separates two classes.
Predicts a number by fitting a straight line to data.
Meta's open family of large language models.
Predicts a yes/no outcome using a logistic curve.
A recurrent network cell that keeps long-term memory.
A measure of how wrong a model's predictions are.
M
Systems that learn patterns from data without explicit programming.
Automatically translating text between languages.
False information spread without intent to deceive.
A model that routes inputs to specialist sub-models.
A platform for tracking experiments and managing models.
Practices for deploying and maintaining machine learning systems reliably.
A trained system that makes predictions from new input data.
A document describing a model's purpose, limits, and risks.
Reducing a model's size while keeping its accuracy.
Putting a trained model into production for real users.
Compressing a large model into a smaller one.
When a live model's accuracy falls as data changes.
Watching a live model for errors and performance drops.
A store that tracks and versions trained models.
Making a trained model available for live requests.
Tracking changes to a model over time.
Several AI agents working together toward shared goals.
AI that works across text, images, audio, and video together.
N
A sequence of N words treated as one unit.
Classifies using probabilities, assuming features act independently.
Finding names of people, places, and organisations in text.
AI built to do one specific task well.
The part of NLP focused on producing human-like text.
AI that lets computers understand and generate human language.
The part of NLP focused on grasping meaning.
Layers of connected nodes that learn patterns from data.
A single unit that takes inputs and produces an output.
Near-term quantum computers with limited, noisy qubits.
A Python library for fast numerical arrays.
O
Finding and locating objects inside an image.
An open format for sharing models across frameworks.
A formal map of concepts and their relationships in a domain.
The company behind GPT models and ChatGPT.
A library for computer vision tasks in code.
Reading text that appears inside images.
Tracking how pixels move between video frames.
An algorithm that updates weights to reduce loss.
When a model memorises training data and fails on new data.
P
A Python library for working with tables of data.
Labelling each word by its grammatical role.
A fingerprint that stays similar for similar-looking images.
A layer that shrinks feature maps by summarising regions.
Finding the positions of body joints in an image.
Encryption designed to survive quantum computers.
Initial broad training before adapting a model to a task.
The share of predicted positives that are actually correct.
Reduces dimensions while keeping most variation.
Protecting personal data from misuse by AI systems.
The instruction or question given to an AI model.
Crafting inputs to get better outputs from a model.
Hidden instructions that hijack a model's behaviour.
Meta's open-source library for building neural networks.
Q
Shrinking a model by using lower-precision numbers.
When a quantum computer solves a useful problem faster.
Solving optimisation problems with quantum tunnelling.
A sequence of gates applied to qubits.
Computing with quantum bits to solve certain problems faster.
Methods that protect qubits from noise and faults.
An operation that changes a qubit's state.
Machine learning run on quantum computers.
When a quantum computer beats the best classical one.
The basic unit of quantum information, holding multiple states.
Systems that answer questions from text or knowledge.
R
Many decision trees voting together for a stronger prediction.
An agent that acts to achieve the best expected outcome.
The share of actual positives a model correctly finds.
A network that remembers earlier inputs in sequences.
Testing a system by attacking it to find weaknesses.
Predicting a continuous number from input data.
Techniques that stop a model from overfitting its training data.
Learning by trial, error, and rewards from actions.
An activation that outputs zero below zero and input above.
A deep network using skip connections to train very deep models.
Building and using AI in a fair, safe way.
Grounding answers by retrieving documents before generating.
Searching the web using an image as the query.
Regression that shrinks coefficients to reduce overfitting.
Tuning a model with human-rated responses.
A score of how well a model separates two classes.
S
A Python library for classic machine learning algorithms.
Attention that lets a sequence weigh its own parts.
Learning by predicting hidden parts of the data itself.
Labelling every pixel by what it shows.
Learning from a small labelled set plus much unlabelled data.
Detecting the feeling or opinion expressed in text.
Running a new model silently beside the live one.
An activation that squeezes values between zero and one.
A function turning scores into probabilities that sum to one.
Converting spoken words into written text.
An open model that generates images from text.
Combining models by training another model on their outputs.
Cutting words to their root by removing endings.
Optimisation using small random data batches.
Common words often removed before text analysis.
AI with genuine understanding and consciousness like humans.
A qubit holding several states at once.
Learning from labelled examples to predict new outcomes.
Finds the best boundary separating classes in data.
Group behaviour from many simple agents following local rules.
AI that reasons with explicit rules, logic, and symbols.
Artificial data created to train or test models.
Hidden instructions that set a model's behaviour.
T
A method that maps high-dimensional data for visualisation.
A hypothetical point when AI surpasses human intelligence.
A setting that controls how random a model's output is.
Google's open-source library for building machine learning models.
Held-out data used to measure final model performance.
Sorting text into predefined categories.
Shortening a long text while keeping key points.
A score of how important a word is in a document.
How many requests a system handles per unit of time.
A small text unit, often a word part, that models process.
Splitting text into tokens for a model to read.
Google's chip built specially for machine learning.
Dividing data into training and testing parts.
Teaching a model by showing it example data.
The examples a model learns from.
Reusing a trained model for a new, related task.
A neural architecture built on attention, behind modern LLMs.
Openness about how an AI system works and decides.
A test of whether a machine can imitate human conversation convincingly.
U
A network shaped like a U for precise image segmentation.
When a model is too simple to capture the data patterns.
Finding hidden patterns in data without labelled examples.
V
Held-out data used to tune a model during training.
When gradients shrink and early layers stop learning.
A generative model learning a smooth latent space.
A database storing embeddings for fast similarity search.
Answering questions about an image.
How alike two images look in content and style.
W
AI that simulates intelligence without real understanding.
A platform for tracking ML experiments.
OpenAI's model that turns speech into text.
A vector that captures a word's meaning.
A method that learns word vectors from surrounding words.
X
A fast, regularised gradient boosting library for tabular data.
Y
A fast object detection system that sees an image once.
Z
Handling a task with no examples at all.