Toward a Unified Framework for Debugging Concept-based Models

TL;DR

This paper proposes a unified framework for debugging concept-based models by addressing bugs in concepts and aggregation functions, introducing a schema for supervision, and a new loss function for robust aggregation.

Contribution

It introduces a schema for human supervision to identify and fix bugs in CBMs and a novel loss function to improve aggregation robustness during training.

Findings

Schema effectively guides bug prioritization in CBMs

New loss function enhances robustness of concept aggregation

Framework generalizes existing alignment strategies

Abstract

In this paper, we tackle interactive debugging of "gray-box" concept-based models (CBMs). These models learn task-relevant concepts appearing in the inputs and then compute a prediction by aggregating the concept activations. Our work stems from the observation that in CBMs both the concepts and the aggregation function can be affected by different kinds of bugs, and that fixing these bugs requires different kinds of corrective supervision. To this end, we introduce a simple schema for human supervisors to identify and prioritize bugs in both components, and discuss solution strategies and open problems. We also introduce a novel loss function for debugging the aggregation step that generalizes existing strategies for aligning black-box models to CBMs by making them robust to how the concepts change during training.