EXIN BCS ML Award key terms

A machine learning approach where a model is trained on labelled example data, learning a mapping from inputs to known outputs so it can predict the label for new, unseen inputs.

A machine learning approach that works with unlabelled data, seeking hidden structure such as groups or compressed representations without any target answer being provided during training.

A learning paradigm where an agent learns by interacting with an environment, taking actions and receiving rewards or penalties, aiming to find a policy that maximises cumulative reward over time.

A hybrid approach that trains on a small set of labelled examples together with a much larger set of unlabelled data, often improving accuracy when labelling is expensive or slow.

The known target value or correct answer attached to a training example, such as a class name or a numeric outcome, which the model tries to reproduce during supervised training.

An individual measurable input variable describing an example, such as age, pixel value, or word count, which the model uses to make its predictions.

The human-driven process of creating, transforming, or selecting input variables so that patterns become easier for a model to learn, often improving performance more than changing the algorithm.

The portion of available data used to fit a model's parameters, providing the examples from which the algorithm learns patterns and adjusts its internal weights.

A held-out portion of data used during development to tune hyperparameters and compare model variants, giving an estimate of performance without touching the final test data.

A portion of data kept completely separate from training and tuning, used once at the end to give an unbiased estimate of how the final model will perform on new data.

When a model learns noise and idiosyncrasies of the training data so closely that it performs well on seen examples but poorly on new, unseen data.

When a model is too simple or insufficiently trained to capture the underlying pattern in the data, resulting in poor performance on both training and new data.

A model's ability to make accurate predictions on data it has never encountered, which is the central goal of all machine learning and the reason held-out data is used.

The balance between a model that is too simple and consistently wrong (high bias) and one that is too sensitive to training data fluctuations (high variance), both harming predictive accuracy.

A supervised task predicting which discrete category an example belongs to, such as spam versus not spam, or which of several disease classes a scan indicates.

A supervised task predicting a continuous numeric value, such as a house price or temperature, rather than a discrete class label.

An unsupervised technique that groups examples so that items within a group are more similar to each other than to items in other groups, without predefined labels.

A supervised model that predicts a continuous output as a weighted sum of input features, fitting a straight line or hyperplane that minimises the error between predictions and actual values.

A classification model that applies a sigmoid function to a weighted sum of features, producing a probability between zero and one that an example belongs to a class.

A model that repeatedly splits data on feature thresholds, forming a tree of if-then rules that leads from input features to a predicted class or value.

An ensemble method that trains many decision trees on random subsets of data and features, then combines their predictions by voting or averaging to reduce variance and overfitting.

A clustering algorithm that partitions data into k groups by repeatedly assigning each point to its nearest centroid and recomputing centroids until assignments stabilise.

A simple instance-based algorithm that classifies or predicts a value for a new example by looking at the k most similar training examples and combining their labels.

A classification model that finds the decision boundary giving the widest margin between classes, using kernel functions to handle non-linear separations.

A model built from layers of interconnected units that each compute a weighted sum of inputs passed through a non-linear activation function, learned by backpropagation.

The algorithm that computes how much each network weight contributed to the output error, working backwards through the layers so that gradient descent can update the weights.

An optimisation method that iteratively adjusts model parameters in the direction that reduces the loss function, stepping against the gradient until the error is minimised.

A hyperparameter controlling how large a step the optimiser takes during each parameter update; too large causes instability, too small makes training very slow.

A mathematical measure of how far a model's predictions are from the true values, which training aims to minimise and which guides parameter updates.

A setting chosen before training rather than learned from data, such as the number of trees in a forest or the learning rate, typically tuned using a validation set.

A technique that penalises model complexity, for example by adding a penalty on large weights to the loss, to reduce overfitting and improve generalisation.

A resampling method that splits data into several folds, training on some and validating on the rest in turn, giving a more reliable performance estimate than a single split.

A table comparing predicted and actual classes, showing true positives, false positives, false negatives, and true negatives, from which metrics like precision and recall are calculated.

The proportion of examples a model labelled as positive that truly are positive; high precision means few false alarms.

The proportion of all actual positive examples that the model correctly identified; high recall means few positive cases are missed.

The harmonic mean of precision and recall, giving a single balanced metric that is useful when classes are imbalanced and both false positives and false negatives matter.

A regression metric that averages the squared differences between predicted and actual values, penalising large errors more heavily than small ones.

Scaling input features to a common range, such as zero to one, so that no variable dominates distance or gradient calculations purely because of its units.

Transforming features so they have zero mean and unit variance, a common preprocessing step that helps many algorithms converge faster and behave more fairly across variables.

Converting a categorical variable into several binary columns, one per category, so that algorithms treat categories as distinct without implying any numeric ordering.

A strategy that combines the predictions of several models, such as by voting or averaging, to produce a result that is usually more accurate and robust than any single model.

An ensemble technique that trains multiple models on different random samples of the training data and averages their outputs, chiefly reducing variance.

An ensemble technique that trains models sequentially, each new one focusing on the errors of the previous ones, typically reducing bias and producing strong predictive accuracy.

Techniques that reduce the number of input variables while retaining most meaningful information, making data easier to visualise and models faster and less prone to overfitting.

A dimensionality reduction method that finds new uncorrelated axes, ordered by how much variance they explain, and projects data onto the most informative few.

The field concerned with enabling computers to process and analyse human language, covering tasks such as text classification, translation, and sentiment analysis.

The field focused on extracting information from images and video, with common tasks including image classification, object detection, and segmentation using pixel-based features.

A modelling error where information from outside the training data, such as test-set statistics, influences training, producing unrealistically good results that fail in production.

The process of integrating a trained model into a live system so its predictions can be used, including concerns such as latency, monitoring, and version control.

The gradual decline in a deployed model's performance as real-world data changes away from the patterns seen during training, requiring monitoring and periodic retraining.

Systematic unfairness in model outcomes, often arising from unrepresentative training data or flawed labelling, which can disadvantage particular groups of people.

The systematic assessment of a model's performance using appropriate metrics, held-out data, and validation procedures, to judge whether it is fit for its intended purpose.