Concept Activation Vector Explainer
A concept activation vector represents a human-defined idea as a direction in a model’s internal activation space. Researchers first collect examples of the concept, such as striped texture, along with comparison examples.
A simple classifier separates the two groups using activations from a chosen layer. The separating direction becomes the concept activation vector, and directional derivatives test whether moving toward that concept increases a target output.
TCAV, a method built around these vectors, can summarize how often a class prediction is sensitive to the concept. It lets people ask questions at a higher level than individual pixels or raw neurons.
The result depends on concept examples, comparison sets, model layer, classifier, and statistical testing. A vector can capture unwanted correlations in the examples, so concept quality and repeated controls are essential.
A concept activation vector represents a human-defined idea as a direction inside a model's activation space and tests its influence on predictions.
Key terms include concept, activation space, direction, layer, comparison set, linear classifier, directional derivative, TCAV score and significance test.
Activations for concept examples are separated from comparison examples.
The separating direction becomes a reusable representation of the concept.
Biased or narrow examples may encode a shortcut instead of the intended idea.
Multiple control sets and significance tests help reveal unstable conclusions.
Explore more "Explainers"
Discover additional explainers across politics, science, business, technology, and other fields. Each explainer breaks down a complex idea into clear, everyday language—helping you better understand how major concepts, systems, and debates shape the world around us.