Computer Science terms in Machine learning

The obligation of the people and organisations who build and deploy a system to take responsibility for its decisions and their consequences, to explain them, and to put right any harm caused.

The proportion of all predictions that a classifier gets right: (TP + TN) divided by the total number of examples.

One of the choices available to an agent in a state, which changes the environment and leads to a reward and a new state.

The function a neuron applies to its weighted sum to produce its output; non-linear activation functions let networks model non-linear patterns.

An input deliberately altered, often by changes too small for a person to notice, so that a machine learning model gives a wrong output.

In reinforcement learning, the program that observes its environment, chooses actions and learns from the rewards those actions earn.

A systematic and repeatable error in a computer system that produces unfair outcomes, such as favouring or disadvantaging one group of people over another.

The property of an automated decision system whose outcomes or error rates do not unjustly differ between groups of people, measured against a stated definition of fairness.

The condition of acting or communicating without one's identity being known; pseudonymity is acting under a consistent false name that is not linked to one's real identity.

The "if" part of an association rule, the itemset on the left of the arrow.

A chip designed and manufactured to perform one specific task, giving the highest speed and energy efficiency for that task but unable to be changed for other tasks.

An algorithm for association rule mining that finds frequent itemsets level by level (single items, then pairs, then triples), counting only candidate itemsets whose subsets are all frequent, then builds rules from them that meet a minimum confidence.

The field of computer science that builds systems able to perform tasks normally thought to need human intelligence, such as recognising speech, making decisions or playing games.

The European Union regulation, in force from 1 August 2024 with its provisions phased in over the following 6 to 36 months, that regulates AI systems according to their level of risk.

A machine learning model made of layers of connected units (artificial neurons), in which each unit combines its weighted inputs and passes the result to the next layer; training adjusts the weights.

A rule of the form X → Y ("if X, then Y") stating that when the items or events in X occur in a record, the items or events in Y also tend to occur, together with measures of how often and how reliably this holds.

Technology that overlays computer-generated images, sound or information onto a person's view of the real world in real time, through a phone screen, headset or smart glasses.

The algorithm that calculates how much each weight in a neural network contributed to the loss, by passing the error backwards from the output layer through the hidden layers.

A learnable number added to a neuron's weighted sum, which shifts the threshold at which the neuron activates, independently of its inputs.

A systematic test of an automated decision system, often by an independent party, that compares its outputs and error rates across groups of people to find and measure unfair differences.

A system whose inputs and outputs can be observed but whose internal reasoning cannot practically be inspected or understood, either because it is too complex or because it is kept secret.

A diagram of one numerical variable that shows its median, lower and upper quartiles as a box, the spread of the remaining non-outlying values as whiskers, and any outliers as separate points.

The centre of a cluster, calculated as the mean of each feature over all the points in the cluster.

In a genetic algorithm, the encoding of one candidate solution.

A situation in which one class is far more common than another in a data set, so that a model can score high accuracy while performing badly on the rare class.

A supervised learning task in which the model predicts which of a fixed set of categories (classes) an input belongs to.

A service from a cloud provider that rents out machine learning hardware such as GPUs and TPUs on demand, together with storage and tools for training, deploying and monitoring models.

An unsupervised learning task that divides unlabelled examples into groups (clusters) so that examples in the same group are more similar to each other than to examples in other groups.

A measure of how well a regression model fits its data, equal to the proportion of the variation in the dependent variable that the model explains, from 0 (none) to 1 (all).

A recommendation technique that predicts a user's interest in an item from the ratings or behaviour of other users with similar tastes.

For a rule X → Y, the proportion of the transactions containing X that also contain Y: support of X and Y together divided by support of X.

A table that counts a classifier's predictions against the actual classes, so that each cell shows how many examples of one actual class were predicted as each class.

A person's free, specific and informed agreement to their data being collected and used for a stated purpose, which they can also withdraw.

The "then" part of an association rule, the itemset on the right of the arrow, which tends to occur when the antecedent does.

The process by which an online platform reviews user-generated content and removes, restricts or labels material that breaks its rules or the law, using automated classifiers, human moderators or both.

The point at which a repeated learning process stops changing its result.

A layer of a CNN that slides each of its kernels across its input, calculating a weighted sum at every position, to produce one feature map per kernel.

A deep neural network designed for images and other grid-like data, which uses convolutional layers of small learnable kernels and pooling layers to learn features from local regions, followed by fully connected layers that classify.

A number from −1 to +1 that measures the strength and direction of the straight-line relationship between two variables.

A method of evaluating a model in which the data is split into k equal parts (folds) and the model is trained k times, each time on k − 1 folds and validated on the remaining fold, with the k scores averaged.

In a genetic algorithm, the operation that combines parts of two parent chromosomes to produce one or more child chromosomes.

The total reward an agent collects over a sequence of actions, which it tries to maximise.

The collection of problems that arise as the number of features grows, including sparse data, the need for exponentially more records, distances between records becoming similar, overfitting, and rising computation and memory costs.

Dividing a customer base into groups whose members behave similarly, so that each group can be targeted differently.

The process of detecting and correcting or removing errors, inconsistencies, duplicates, irrelevant entries and gaps in a dataset so that it accurately represents what it describes.

The use, during training or preprocessing, of information that would not be available when the model is used for real, which makes the model's measured performance unrealistically high.

The principle that an organisation should collect and keep only the personal data that is adequate, relevant and necessary for its stated purpose.

An attack in which false or manipulated examples are added to a model's training data so that the trained model behaves wrongly, either in general or for inputs chosen by the attacker.

The set of steps that turn raw data into a form a machine learning algorithm can learn from, including cleaning, encoding, scaling and selecting features.

The degree to which a dataset is accurate, complete, consistent, valid, unique, relevant and up to date for the purpose it is used for.

The condition in which most of the possible combinations of feature values have no records, either because the space is too large for the data to fill or because most values in the dataset are empty or zero.

A structured collection of examples used to train or test a machine learning model, where each example records the same set of properties.

A supervised learning model that classifies data by following a path from a root node through tests on feature values, one test per internal node, to a leaf node that gives the predicted class.

Machine learning that uses artificial neural networks with many layers, which learn their own features in stages directly from raw data such as images, audio or text.

An image, video or audio recording of a real person that has been created or altered using deep learning so that it appears authentic.

Handling missing data by removing every record that contains a missing value, or by removing a feature whose values are mostly missing.

The output variable in a model, whose value is predicted from the independent variable or variables; also called the response.

The gap between people, groups or countries that have access to modern digital technology, connectivity and the skills to use them and those that do not.

The number of features used to describe each record in a dataset, which is the number of dimensions of the space its records lie in.

Reducing the number of features in a dataset while preserving the information that matters for the task, by selecting a subset of features or by combining features into fewer new ones.

False information that is created or spread deliberately to deceive, for example to influence an election, damage a reputation or make money.

A rule for calculating how far apart two records are from their feature values, such as Euclidean or Manhattan distance, used by algorithms that compare records by similarity.

A record that describes the same real-world entity or event as another record in the dataset, either exactly or with small differences in spelling or formatting.

A device at the edge of a network, close to where data is produced, such as a phone, camera, sensor or car, that runs machine learning models locally instead of sending data to a remote server.

A way of choosing the number of clusters k by plotting the within-cluster sum of squares against k and picking the value at the "elbow", where further increases in k give only small reductions.

Discarded electrical and electronic equipment, such as computers, phones and servers, which contains valuable and toxic materials and is often exported to countries with weaker recycling controls.

A feature selection strategy in which the learning algorithm itself selects or down-weights features while the model is being trained.

A model that combines the predictions of several different models in order to predict more accurately than any of them alone.

Everything outside the agent that it interacts with: it receives the agent's actions and responds with a new situation and a reward.

The effect of a technology on the natural environment over its life cycle, including the energy and water used to make and run it, the greenhouse gases emitted and the waste produced when it is discarded.

One complete pass of the whole training data set through a learning algorithm during training.

An action-selection rule in which an agent explores by choosing a random action with probability ε and otherwise exploits by choosing the action with the highest estimated value.

Fairness in how the benefits, costs and risks of a technology are distributed, taking into account that different groups start from different positions and may need different support to benefit equally.

The study of what is right and wrong in human conduct, applied in computing to the choices made when designing, building and using technology and to their effects on people and the environment.

The straight-line distance between two points, the square root of the sum of the squared differences between their coordinates.

The degree to which the reasons for a model's individual outputs can be described in terms a person can understand, such as which inputs most influenced a decision.

In reinforcement learning, choosing the action with the highest estimated reward based on what the agent has learned so far.

In reinforcement learning, choosing an action that is not currently believed to be the best, in order to gather information about its reward.

Using a model to predict values outside the range of the data it was fitted to, as opposed to interpolation, which predicts within that range.

The harmonic mean of precision and recall, 2 × precision × recall ÷ (precision + recall), which is high only when both are high.

The use of computer vision, usually a deep learning model, to detect a face in an image and identify or verify a person by comparing the face with stored images.

An example that is actually positive but that the model wrongly predicts as negative; a missed case.

An example that is actually negative but that the model wrongly predicts as positive; a false alarm.

One measurable property of an example that is given to a machine learning model as an input, such as the number of links in an email or the floor area of a house.

Creating new features by combining or transforming existing ones, so that fewer or more informative features describe each record.

A score for each feature showing how much it contributes to a trained model's predictions.

The grid of values a convolutional layer produces by applying one kernel across its input, showing how strongly the kernel's pattern is present at each position.

Transforming the values of numerical features so that they lie on comparable scales, so that no feature dominates a model because of its units or range.

The process of choosing a subset of the existing features of a dataset that are most useful for predicting the target, and discarding the others.

A cycle in which a model's outputs influence the data later used to retrain or evaluate it, so that its errors or biases are reinforced over time.

A chip containing a large grid of logic blocks whose connections can be reconfigured after manufacture, so it can be programmed to act as a custom hardware circuit for a particular task and reprogrammed later.

The narrowing of the information a person sees online when personalisation algorithms keep showing content that matches their past behaviour and views, reducing exposure to other perspectives.

A feature selection strategy that ranks features by a statistical measure of their relationship with the target, such as correlation, calculated without training any model, and keeps the highest-ranked features.

Further training of a pre-trained model on a smaller, task-specific dataset, usually with a new output layer and small changes to the existing weights.

A function that gives each candidate solution in a genetic algorithm a score showing how well it solves the problem.

Floating-point operations per second, a measure of a processor's arithmetic speed; large values use prefixes such as tera (${10}^{12}$), peta (${10}^{15}$) and exa (${10}^{18}$).

A layer of a neural network in which every neuron receives an input from every neuron in the previous layer, each connection having its own weight.

The principle that the output of a computer system can be no better than its input, so a model trained on poor-quality data produces poor-quality predictions.

The European Union law, applied since 25 May 2018, that sets rules for processing personal data, gives individuals rights over their data, and allows fines of up to €20 million or 4% of a company's worldwide annual turnover, whichever is higher.

The ability of a trained model to make accurate predictions on new data that it did not see during training.

One complete cycle of a genetic algorithm (evaluation, selection, crossover and mutation), or the population that results from it.

Machine learning models that produce new content, such as text, images, audio or code, resembling the data they were trained on.

A search and optimisation method, inspired by natural selection, that improves a population of candidate solutions over many generations through selection of the fittest, crossover and mutation.

A measure of how mixed the classes in a node are, equal to the probability that two examples drawn at random from the node (with replacement) belong to different classes; 0 for a pure node.

An optimisation method that repeatedly changes a model's parameters by a small step in the direction that reduces the loss most steeply.

A layer of neurons between the input and output layers of a neural network, which transforms the outputs of the previous layer into new features.

The use of supercomputers, very large clusters of processors connected by fast networks and run as one system, to solve problems too large for ordinary computers.

A design in which a person reviews, approves or can override an automated system's output before it takes effect, especially for decisions with serious consequences.

A setting chosen before training that controls how a model learns or how complex it is, and that is not learned from the training data.

The process of searching for the hyperparameter values that give a model the best performance on validation data.

Values that are recorded in inconsistent or unusable forms, such as mixed date formats, mixed units, inconsistent spelling or capitals, or numbers stored as text.

Handling missing data by replacing each missing value with an estimate, such as the mean, median or mode of the feature, or the previous value in a time series.

Values that are present but do not match reality, such as impossible values, values that break a rule of their field, or values that contradict other fields in the same record.

The input variable in a model, whose values are used to predict another variable; also called the predictor.

The freedoms and protections each person is entitled to, such as privacy, freedom of expression, equal treatment and due process, including how these apply to digital technology and data.

The use of a trained model to produce an output, such as a prediction or classification, for new input data that the model has not seen before.

The reduction in entropy (a measure of how mixed the classes are) achieved by splitting a node on a feature; the split with the highest information gain is chosen.

The first layer of a neural network, which receives the input features, one neuron per feature, and passes them to the next layer.

In a regression equation, the predicted value of the dependent variable when the independent variable is zero.

The degree to which a person can understand how a model reaches its predictions.

The difference between the upper quartile and the lower quartile of a set of values, $Q_3-Q_1$, which measures the spread of the middle 50% of the data.

Records outside the population a model is meant for, or fields that carry no useful information about what the model must predict.

A set of one or more items that may occur together in a record.

An unsupervised learning algorithm that divides data into k clusters by repeatedly assigning each point to its nearest centroid and then moving each centroid to the mean of the points assigned to it.

A supervised learning algorithm that classifies a new data point by finding the k training points closest to it and assigning the class that most of them share.

A small grid of learnable weights that a convolutional layer slides across its input to detect one kind of local pattern.

The known correct output attached to an example in a dataset, which a supervised learning model learns to predict from the features.

A very large deep learning model, pre-trained on huge amounts of text, that processes and generates human language by repeatedly predicting the next part of a text.

A graph of a model's error (or accuracy) on training data and on validation data plotted against the amount of training, the amount of data or the model's complexity, used to detect underfitting and overfitting.

A hyperparameter that sets the size of the change made to a model's weights at each step of training.

The method of fitting a regression line by choosing the slope and intercept that minimise the sum of the squared residuals.

For a rule X → Y, the confidence of the rule divided by the support of Y; a lift above 1 means X and Y occur together more often than if they were independent, 1 means no association, and below 1 means a negative association.

A supervised learning method that models the relationship between one or more input variables and a continuous output variable with a straight line (or flat surface), and uses that line to predict the output for new inputs.

Describes two classes of data that can be completely separated by a straight line (or, in more dimensions, a flat plane).

A function that measures how far a model's predictions are from the correct values, giving a single number that training tries to make as small as possible.

A branch of artificial intelligence in which a program improves its performance at a task by finding patterns in data, instead of following rules written explicitly by a programmer.

A rule that assigns the class held by the largest number of the votes being counted.

The distance between two points measured along axes at right angles, equal to the sum of the absolute differences between their coordinates.

The analysis of customers' purchase records to find products that are frequently bought together.

The use of machine learning to analyse medical images, such as X-rays, scans or photographs of skin, to detect or classify signs of disease.

False or misleading information that is shared without the intention to deceive, often by people who believe it is true.

Values that should be present in a dataset but were not recorded, shown as blanks, NULL, None or a placeholder such as −1 or 999.

The result of training a machine learning algorithm on data: a stored set of learned values and structure that maps an input to an output, such as a prediction or a decision.

Techniques that reduce the size and computing needs of a trained model so it can run on limited hardware, such as storing weights with fewer bits (quantisation) or removing weights that contribute little (pruning).

A value that a model learns from its training data during training.

The process of choosing, from several candidate algorithms or trained models, the one best suited to a problem, by considering the nature of the task and data and comparing performance fairly on the same validation data.

A feedforward neural network made of an input layer, one or more hidden layers and an output layer, in which each layer is fully connected to the next.

In a genetic algorithm, a small random change to a chromosome, used to maintain diversity in the population.

The branch of AI that enables computers to analyse, understand and generate human language in text or speech.

In an artificial neural network, a processing unit that calculates a weighted sum of its inputs plus a bias and passes the result through an activation function to produce its output.

The result that no learning or optimisation algorithm performs best on every possible problem, so an algorithm's advantage on some problems is balanced by worse performance on others.

Rescaling a feature to a fixed range, usually 0 to 1, using $x^{\mathrm{\prime }}={\displaystyle \frac{x-x_{\min }}{x_{\max }-x_{\min }}}$, where $x_{\min }$ and $x_{\max }$ are the smallest and largest values in the training data.

The machine learning task of finding objects in an image or video, marking each one's position (usually with a bounding box) and classifying what it is.

A way of turning a categorical feature into numbers by creating one new binary column per category, set to 1 for the record's category and 0 for all the others.

Repeated or severe abusive behaviour directed at a person through digital channels, such as threats, insults, stalking, doxxing (publishing private details) or sharing sexual images without consent.

A problem in which the aim is to find the best solution from a set of possible solutions, according to a measurable objective.

A data value that lies an abnormally long way from the other values of the same feature.

The final layer of a neural network, whose neurons produce the network's predictions.

The result of a model learning the noise and specific details of its training data instead of the general pattern, so that it performs well on the training data but poorly on new data.

Extra pixels, usually zeros, added around the border of a CNN layer's input so that kernels can be applied at the edges and the output keeps its size.

The simplest artificial neural network, a single neuron that calculates a weighted sum of its inputs plus a bias and applies a step activation function to output one of two classes.

The embedding of artificial intelligence throughout everyday objects, services and institutions, such as phones, homes, cars, workplaces and public services, so that people interact with AI systems constantly, often without noticing.

The strategy an agent follows, which maps each state to the action (or the probabilities of the actions) it will take.

A CNN layer that reduces the width and height of each feature map by replacing each small window of values with a single summary value.

In a genetic algorithm, the set of candidate solutions (chromosomes) that exist at one stage of the search.

Public-key cryptographic algorithms designed to remain secure against attacks by both classical and quantum computers, intended to replace algorithms such as RSA that a large quantum computer could break.

A model that has already been trained on a large, general dataset and is made available for reuse as the starting point for other tasks.

The proportion of a classifier's positive predictions that are actually positive: TP divided by (TP + FP).

Handling missing data by training a model on the complete records to predict the missing feature from the other features, then using its predictions to fill the gaps.

The right of individuals to control the collection, use and sharing of information about themselves, and to be free from unwanted observation.

A feature that is not itself a protected attribute but is closely correlated with one, so that a model using it can treat groups differently without being given the protected attribute directly.

Reducing the size of a decision tree, by limiting its growth or by removing branches after it is built, so that it captures general patterns rather than noise in the training data.

Computing that uses quantum-mechanical effects such as superposition and entanglement, with quantum bits (qubits) as the basic unit of information, to solve certain problems far faster than classical computers can.

The proportion of actual positive examples that a classifier correctly identifies: TP divided by (TP + FN); also called sensitivity.

The region of the input image that can affect the value of a particular neuron in a CNN; it grows larger in deeper layers.

A supervised learning task in which the model predicts a continuous numerical value, such as a price, a temperature or a score.

A type of machine learning in which an agent learns by trial and error which actions to take in an environment, by receiving rewards for good outcomes and penalties for bad ones.

The difference between the actual value of the dependent variable for a data point and the value predicted by the model: residual = actual − predicted.

A numerical signal that the environment gives an agent after an action, positive for desirable outcomes and negative (a penalty) for undesirable ones, which the agent learns to maximise over time.

The use of machine learning and sensor data to let a robot work out where it is and plan and control its movement through an environment towards a goal while avoiding obstacles.

The step that chooses chromosomes to act as parents for the next generation, giving fitter chromosomes a greater chance of being chosen.

The use of natural language processing to classify the opinion or emotion expressed in a piece of text, for example as positive, negative or neutral.

In a regression equation, the change in the predicted value of the dependent variable for each increase of one unit in the independent variable.

The effect of a technology on society as a whole, including on jobs, inequality, relationships between groups, the distribution of power and how people live and make decisions.

The arrangement of features learned by the layers of a CNN, from simple local patterns in early layers to complex, larger-scale structures in deeper layers.

The measure a decision tree algorithm uses to choose which feature (and threshold) to test at a node, by scoring how well each possible split separates the classes.

Any person, group or organisation that affects, or is affected by, a computer system, including its developers, the organisation that deploys it, its users, the people it makes decisions about, and wider society.

Rescaling a feature so that it has a mean of 0 and a standard deviation of 1, by replacing each value $x$ with its z-score ${\displaystyle \frac{x-\mu }{\sigma }}$, using the mean $\mu$ and standard deviation $\sigma$ of the training data.

A description of the situation the agent is in at a given time, containing the information it uses to choose its next action.

The number of pixels by which a kernel (or pooling window) moves between one position and the next.

A type of machine learning in which a model is trained on labelled data, learning to map each example's features to its known correct output so that it can predict the output for new examples.

The proportion of all transactions that contain every item in an itemset; for a rule X → Y, the proportion containing both X and Y.

The systematic monitoring of people's behaviour, communications or movements, by governments or organisations, to gather information or influence and control them.

An application-specific integrated circuit designed by Google to accelerate the matrix calculations of neural networks, used for both training and inference.

The rule that determines when a genetic algorithm stops.

The part of a labelled dataset kept back from training and used afterwards to measure how accurately the model performs on examples it has not seen.

The stage in which a machine learning algorithm repeatedly adjusts a model's internal values using example data, so that the model's outputs become more accurate.

The part of a labelled dataset used to train a machine learning model.

Bias in a model that comes from the data it learned from, because some groups are under-represented (sampling bias), the data records past unfair decisions (historical bias), or the labels reflect the judgements or mistakes of the people who assigned them (labelling bias).

A machine learning approach in which a model trained for one task is reused as the starting point for a related task, so that the new task needs far less data and training time.

Openness about how an automated system works, including what data it uses, how it was trained and tested, where it is used, and its known limitations.

The problem of finding the shortest route that visits each of a set of locations exactly once and returns to the starting point.

An example that is actually negative and that the model correctly predicts as negative.

An example that is actually positive and that the model correctly predicts as positive.

The result of a model being too simple to capture the underlying pattern in the data, so that it performs poorly on both the training data and new data.

A type of machine learning in which a model is trained on unlabelled data and finds structure in it, such as groups of similar examples or items that occur together, without being given correct outputs.

A portion of the labelled data kept out of training and used to compare models or hyperparameter values, separate from the test set used for the final evaluation.

Technology that places a person in a fully computer-generated, interactive environment, usually through a headset that tracks the movements of their head and hands.

A number attached to a connection in a neural network, by which the input on that connection is multiplied, so that it controls how strongly that input affects the neuron's output.

The total of each input multiplied by its weight, plus the bias, calculated by a neuron before its activation function is applied.

A feature selection strategy that searches through subsets of features by training and evaluating a model on each subset, and keeps the subset that gives the best performance.

The number of standard deviations a value lies above or below the mean of its feature, calculated as $z={\displaystyle \frac{x-\mu }{\sigma }}$.