Labeling Neural Representations with Inverse Recognition

Kirill Bykov will be presenting a talk on “Labeling Neural Representations with Inverse Recognition” at NeurIPS 2023. The talk will discuss a scalable approach, Inverse Recognition (INVERT), for connecting learned representations with human-understandable concepts by leveraging their capacity to discriminate between these concepts. The approach does not rely on the availability of segmentation masks and provides an interpretable metric assessing the alignment between the representation and its corresponding explanation, delivering a measure of statistical significance.