Explore
Artificial Intelligence

A transformation applied to a neuron’s input sum. Nonlinear activations let networks model nonlinear relationships.

A system that selects actions toward a goal using observations and available tools; the term has several definitions.

A mechanism that computes weighted combinations of representations, often across sequence positions.

An efficient application of the chain rule to compute gradients through a computational graph.

An adjustable offset in a model. This is distinct from unfairness or statistical bias.

Agreement between predicted probabilities and observed frequencies across comparable predictions.

Predicting a category or category probabilities from input data.

Grouping examples using a chosen representation and similarity rule, without supplied cluster labels.

A restriction a solution must satisfy; soft constraints instead express preferences.

The bounded amount of input and generated content a model can consider in one interaction.

A local operation that applies shared filter weights across positions, commonly used in image models.

Use of information during training or evaluation that would not legitimately be available at prediction time.

Machine learning using neural networks with multiple processing layers.

A difference between the data distribution used for development and the one encountered later.

A numerical vector representation whose learned relationships can support tasks such as similarity search.

One pass through the training dataset; it is not a universal unit of model quality.

An input measurement or representation used by a model.

Training that updates some or all parameters of an already trained model for a task or domain.

Systems that produce content such as text, images or audio using learned patterns.

A vector of partial derivatives describing local change in a scalar function.

Connecting an answer to relevant evidence or inputs; it does not guarantee the answer is correct.

Generated content that is unsupported or incorrect despite appearing plausible.

A rule or estimate used to guide a problem-solving process, such as estimated remaining path cost.

Applying a model or reasoning procedure to inputs; in ML this usually uses already fitted parameters.

The target outcome associated with a supervised training example.

A setting controlling the size of optimizer updates during training.

A numerical training objective measuring prediction error or other desired behavior.

A mathematical or computational representation used to produce predictions, content or decisions.

A parameterized network of transformations, typically trained using examples.

Learning training-specific details that do not generalize well to unseen data.

A fitted value inside a model, such as a weight or bias.

In decision-making systems, a rule or learned mapping that selects actions from observations or states.

True positives divided by all predicted positives, when the denominator is nonzero.

An attempt to make a system treat untrusted content as instructions that change its authorized behavior.

Retrieval-augmented generation: retrieving external material and using it as context for generation.

True positives divided by all actual positives, when the denominator is nonzero.

Predicting a numeric outcome, such as demand or travel time.

Techniques that constrain learning to improve generalization rather than merely fit training data.

Learning a policy from reward signals associated with actions and resulting transitions.

A function that maps a vector of scores to nonnegative values summing to one.

Learning from examples paired with target outcomes.

Data reserved for final evaluation after development choices are fixed.

A unit used to represent content in a model; it may be a word fragment, symbol or other encoded element.

A neural architecture based on attention and other components, with variants for different tasks.

Learning structure without supplied target labels for the main learning task.

Data used during development to compare models and select settings.

A defined sequence or graph of processing steps; it can use AI without granting open-ended action selection.