Model & Data Insights using Pre-trained Language Models

Saeid Asgari, Aliasghar Khani, Amir Hosein Khasahmadi, Aditya Sanghi, Karl D.D. Willis, Ali Mahdavi Amiri

Keywords: Data bias, spurious correlations, interpretability

TL;DR: We demonstrate how pretrained language models can be utilized to gain insights into data and understand the features learned by a vision model

Abstract: We propose TExplain, using language models to interpret pre-trained image classifiers' features. Our approach connects the feature space of image classifiers with language models, generating explanatory sentences during inference. By extracting frequent words from such explanations, we gain insights into learned features and patterns. This method detects spurious correlations and biases within a dataset, providing a deeper understanding of the classifier's behavior. Experimental validation on diverse datasets, including ImageNet-9L and Waterbirds, shows potential for improving interpretability and robustness in image classifiers.

Submission Number: 2