Standardized human tests show that most vision models are face-blind

A recent study has revealed that the majority of human face recognition models, 55% to be precise, exhibit a level of performance that places them in the same category as individuals with face-blindness, or prosopagnosia. In contrast, even the most advanced models do not reach the level of super-recognizers, individuals who can accurately identify strangers with exceptional accuracy.

The researchers administered standardized human face recognition tests to a diverse range of models, enabling them to understand where these models fit within the spectrum of human face recognition ability. Upon analyzing the internal representations of the models, they discovered that those that performed well on standardized face recognition tasks demonstrated greater identity-selectivity and viewpoint-invariance.

However, the study also uncovered that models with lower performance still contained some identity information in separate representational subspaces. Removing the viewpoint-dependent subspaces improved the face recognition abilities of 49 out of the 53 models tested. Further scrutiny through targeted unit ablations provided insights into the opposing effects of viewpoint-dependent and viewpoint-tolerant units. Viewpoint-dependent units appeared to hinder identity coding, while viewpoint-tolerant units facilitated it.

Overall, the findings of this research indicate that most AI models currently possess a limited ability to recognize faces, falling significantly short of human performance. This study also highlights the potential for using differences in AI model performance to generate testable hypotheses about the computational mechanisms underlying human face recognition, which can subsequently be investigated in future studies.