المسرد

Computer methods that perform tasks such as pattern recognition, prediction, generation or automation.

A defined process or sequence of steps used to perform a task.

A mathematical or computational representation that transforms inputs into outputs.

A system that follows explicitly defined logic or rules.

Methods in which models learn patterns from data to perform a task.

Learning from examples with known target labels or values.

Finding patterns or structure without predefined target labels.

Machine learning using neural networks with multiple processing layers.

A layered mathematical model that learns relationships by adjusting parameters.

The process of fitting or adjusting a model using data.

Using a trained model to produce an output from new input.

An estimate of an outcome, state or numerical value.

Assigning an input to one or more categories.

AI that creates new content such as text, image or audio.

A model trained on large datasets to process and generate language.

A unit of text processed by a language model.

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

A model that processes or combines more than one data type, such as text and images.

AI methods for processing and interpreting images or video.

Assigning image pixels or regions to structures or categories.

The fitness of data for a task, including accuracy, completeness and relevance.

Information about where data came from and how it was collected or transformed.

Values or observations absent from a dataset or input.

A target category or value associated with an example.

The reference value treated as the correct answer for development or evaluation.

A measurable variable used as a substitute for a concept that is harder to measure directly.

Data used to fit a model.

Data used during development to tune choices or compare models.

Held-out data used for performance assessment.

Improper information flow that gives a model information it should not have during development or evaluation.

Learning development-data detail that does not generalise well to new data.

A model's ability to perform appropriately on relevant new data.

Evaluation using data meaningfully separate from the development setting.

A change between development data and data encountered in real use.

The proportion of people with the target condition who test positive.

The proportion of people without the target condition who test negative.

The proportion of positive results that truly have the target condition.

The proportion of negative results that truly do not have the target condition.

The proportion of a population with a condition in a defined period.

A cut-off used to convert a score or probability into a class or action.

A plausible-looking AI output that is false, unsupported or inconsistent with available evidence.

Creation of false facts, details, data or references.

A workflow that retrieves external content and supplies it to a generative model as context.

Finding and selecting relevant information from a source or knowledge base.

Connecting an output to supplied evidence or sources.

Instructions and context supplied to a generative AI model.

A result produced in a predictable schema or set of fields.

A defined structure describing fields and expected data types.

Meaningful human review, judgement or control over an AI-supported process.

The tendency to over-trust suggestions from an automated system.

Over-reliance on an initial piece of information or early hypothesis.

Systematic patterns that can produce distorted or unfair outcomes.

Consideration of equitable treatment and outcomes across people and groups.

Performance measured separately for relevant patient or user groups.

The extent to which relevant reasons or factors related to an AI output can be communicated or understood.

Communication of relevant information about an AI system, its purpose, limits and use.

Incomplete confidence about a fact, estimate, diagnosis or model output.

Using only personal data that is adequate, relevant and necessary for the purpose.

Reducing or removing information that directly or indirectly identifies a person.

Linking supposedly de-identified data back to an identifiable person.

Malicious or untrusted instructions designed to manipulate an AI system's behaviour.

Giving a user or system only the permissions needed for the task.

An AI system that can select steps, use tools or take actions toward a goal.

Tools for building applications or workflows mainly through visual or configured components.

An early version built to explore a concept or design.

A limited real-world test before wider deployment.

A system deployed for routine real-world use.

The purpose, users, population and context for which a product is designed.

Software intended for a medical purpose that performs that purpose without being part of a hardware medical device.

Structures, roles and processes used to direct, control and account for a system or programme.

A controlled process for reviewing, approving and documenting changes.

Structured assessment of evidence that medical software is clinically appropriate for its intended purpose.

Ongoing collection and review of information about performance, safety or use.

The authoritative source against which an AI output should be checked.