Multimodal Concept Activation Vectors (MCAV)

MCAV defines concepts using exemplar datasets that contain both visual and textual representations of an idea. For instance, the concept of 'danger' might include images of warning signs paired with text descriptions of hazardous situations. This dual-modality definition allows the technique to test whether a vision-language model has learned a unified, cross-modal representation of the concept rather than separate, modality-specific features.

At its core, MCAV computes a concept activation vector (CAV) in the model's joint embedding space by training a linear classifier to distinguish concept exemplars from random counterexamples. It then measures the directional derivative of the model's prediction with respect to this vector. A high sensitivity score indicates that perturbing the input toward the concept direction—across both modalities—consistently changes the output, confirming the concept's influence on the model's reasoning.

Unlike unimodal explanation methods, MCAV operates in the shared representation space where visual and textual features are fused. This means a single concept vector can capture semantic meaning regardless of whether the evidence originated in the image or the text. For example, the concept of 'crowdedness' can be activated by either a dense visual scene or a textual description of a packed venue, and MCAV quantifies the model's reliance on this abstracted idea.

MCAV inherits the TCAV score methodology, which reports the fraction of test inputs whose predictions are positively influenced by the concept vector. Statistical significance is established through a two-sided t-test comparing these directional derivatives against a null distribution of random vectors. This provides a rigorous, quantitative measure of concept sensitivity rather than a subjective visual inspection, making it suitable for auditing and compliance workflows.

MCAV can be applied at any layer of a multimodal architecture, but it is particularly powerful when probing fusion layers where cross-modal interactions occur. By testing concept sensitivity before and after fusion, engineers can diagnose whether a concept is learned independently within each modality or emerges from the interaction itself. This helps identify brittle cross-modal correlations that may not generalize.

A primary use case for MCAV is testing for undesirable concept associations in deployed vision-language models. Auditors can define concepts related to protected attributes or harmful stereotypes using multimodal exemplars and then quantify the model's sensitivity to these concepts across diverse inputs. A statistically significant TCAV score for a harmful concept indicates the model has internalized that association, flagging it for mitigation.