Unsupervised Model Diagnosis

TL;DR

This paper introduces UMO, an unsupervised framework that uses generative models to identify and visualize model vulnerabilities and failure modes in vision systems without human guidance.

Contribution

UMO is the first unsupervised approach to diagnose model robustness by discovering semantic counterfactuals without user input, leveraging generative models and text-based attribute matching.

Findings

Successfully identifies spurious correlations in vision models.

Visualizes failure modes without human intervention.

Validates on multiple vision tasks like classification and segmentation.

Abstract

Ensuring model explainability and robustness is essential for reliable deployment of deep vision systems. Current methods for evaluating robustness rely on collecting and annotating extensive test sets. While this is common practice, the process is labor-intensive and expensive with no guarantee of sufficient coverage across attributes of interest. Recently, model diagnosis frameworks have emerged leveraging user inputs (e.g., text) to assess the vulnerability of the model. However, such dependence on human can introduce bias and limitation given the domain knowledge of particular users. This paper proposes Unsupervised Model Diagnosis (UMO), that leverages generative models to produce semantic counterfactual explanations without any user guidance. Given a differentiable computer vision model (i.e., the target model), UMO optimizes for the most counterfactual directions in a generative latent space. Our approach identifies and visualizes changes in semantics, and then matches these changes to attributes from wide-ranging text sources, such as dictionaries or language models. We validate the framework on multiple vision tasks (e.g., classification, segmentation, keypoint detection). Extensive experiments show that our unsupervised discovery of semantic directions can correctly highlight spurious correlations and visualize the failure mode of target models without any human intervention.