Language-assisted Vision Model Debugger: A Sample-Free Approach to Finding Bugs

Vision models with high overall accuracy often exhibit systematic errors in specific scenarios, posing potential serious safety concerns. Diagnosing bugs of vision models is gaining increased attention, however traditional diagnostic approaches require annotation efforts (\eg rich metadata accompanying each samples of CelebA). To address this issue,We propose a language-assisted diagnostic method that uses texts instead of images to diagnose bugs in vision models based on multi-modal models (\eg CLIP). Our approach connects the embedding space of CLIP with the buggy vision model to be diagnosed; meanwhile, utilizing a shared classifier and the cross-modal transferability of embedding space from CLIP, the text-branch of CLIP become a proxy model to find bugs in the buggy model. The proxy model can classify texts paired with images. During the diagnosis, a Large Language Model (LLM) is employed to obtain task-relevant corpora, and this corpora is used to extract keywords. Descriptions constructed with templates containing these keywords serve as input text to probe errors in the proxy model. Finally, we validate the ability to diagnose existing visual models using language on the Waterbirds and CelebA datasets, we can identify bugs comprehensible to human experts, uncovering not only known bugs but also previously unknown ones.

Understanding and Testing Generalization of Deep Networks on Out-of-Distribution Data

Deep network models perform excellently on In-Distribution (ID) data, but can significantly fail on Out-Of-Distribution (OOD) data. While developing methods focus on improving OOD generalization, few attention has been paid to evaluating the capability of models to handle OOD data. This study is devoted to analyzing the problem of experimental ID test and designing OOD test paradigm to accurately evaluate the practical performance. Our analysis is based on an introduced categorization of three types of distribution shifts to generate OOD data. Main observations include: (1) ID test fails in neither reflecting the actual performance of a single model nor comparing between different models under OOD data. (2) The ID test failure can be ascribed to the learned marginal and conditional spurious correlations resulted from the corresponding distribution shifts. Based on this, we propose novel OOD test paradigms to evaluate the generalization capacity of models to unseen data, and discuss how to use OOD test results to find bugs of models to guide model debugging.