Glossary Terms

Methods that predict or group observations using similar examples in a dataset. Results depend on the similarity measure, feature representation and whether the stored examples represent the intended use.

Methods for exploring alternative neural networkA model made of connected computational units with adjustabl structures. A higher benchmark score should be weighed against search cost, operating requirements and performance under the actual application conditions.

A model made of connected computational units with adjustable parameters that learn patterns from data. Its outputs require evaluation for the intended task; the architecture alone does not establish reliability.

Computer visionAI methods for interpreting information in images or video, methods that identify and locate objects within an image or video. Test relevant object types, environments and missed or incorrect detections before relying on the output in operations.

An explicit model of concepts, categories and relationships within a domain. A useful ontologyAn explicit model of concepts, categories and relationships helps teams share meaning and organise information consistently, with definitions maintained as the domain changes.

Consistent improvement in how an organisation delivers value through its processes and capabilities. Assess quality, reliability, customer outcomes and total effort rather than speed alone.

Choices about how operations will support the organisation’s business objectives. It connects capacity, processes, capabilities and systems, including where AI may contribute and where oversight is required.

An algorithm used to update model parameters during training. Select and evaluate it alongside the loss functionThe mathematical objective used to measure model error durin, data and training configuration rather than assuming one method is best for every task.

Using an external provider to perform an activity or service. Define responsibilities, information access, quality requirements and oversight, because accountability is not automatically transferred with the work.

When a model learns the training data too closely and performs poorly on sufficiently different examples. Assess it using data not used for training and representative real-world conditions.