Measuring the Interpretability of Unsupervised Representations via Quantized Reverse Probing

TL;DR

This paper introduces a novel method called reverse linear probing to measure the interpretability of self-supervised visual representations by estimating the mutual information between representations and labeled concepts, revealing semantic content and concept combinations.

Contribution

It proposes a new interpretability metric based on reverse linear probing, capable of detecting semantic and compositional information in representations, and introduces automatic concept labeling for large datasets.

Findings

The method effectively ranks representations by interpretability.

It detects when representations encode concept combinations.

It reveals differences from standard linear probe evaluations.

Abstract

Self-supervised visual representation learning has recently attracted significant research interest. While a common way to evaluate self-supervised representations is through transfer to various downstream tasks, we instead investigate the problem of measuring their interpretability, i.e. understanding the semantics encoded in raw representations. We formulate the latter as estimating the mutual information between the representation and a space of manually labelled concepts. To quantify this we introduce a decoding bottleneck: information must be captured by simple predictors, mapping concepts to clusters in representation space. This approach, which we call reverse linear probing, provides a single number sensitive to the semanticity of the representation. This measure is also able to detect when the representation contains combinations of concepts (e.g., "red apple") instead of just individual attributes ("red" and "apple" independently). Finally, we propose to use supervised classifiers to automatically label large datasets in order to enrich the space of concepts used for probing. We use our method to evaluate a large number of self-supervised representations, ranking them by interpretability, highlight the differences that emerge compared to the standard evaluation with linear probes and discuss several qualitative insights. Code at: {\scriptsize{\url{https://github.com/iro-cp/ssl-qrp}}}.