Tackling Shortcut Learning in Deep Neural Networks: An Iterative Approach with Interpretable Models

TL;DR

This paper introduces an iterative method using interpretable models with First Order Logic to identify and eliminate shortcut learning in deep neural networks, improving interpretability and robustness without sacrificing accuracy.

Contribution

The paper proposes a novel iterative approach combining concept-based interpretable models with residual networks to detect and remove shortcuts in deep models, enhancing interpretability.

Findings

Effective shortcut detection using FOL from interpretable experts

Elimination of shortcuts via finetuning with Metadata Normalization

Maintains original model accuracy while removing biases

Abstract

We use concept-based interpretable models to mitigate shortcut learning. Existing methods lack interpretability. Beginning with a Blackbox, we iteratively carve out a mixture of interpretable experts (MoIE) and a residual network. Each expert explains a subset of data using First Order Logic (FOL). While explaining a sample, the FOL from biased BB-derived MoIE detects the shortcut effectively. Finetuning the BB with Metadata Normalization (MDN) eliminates the shortcut. The FOLs from the finetuned-BB-derived MoIE verify the elimination of the shortcut. Our experiments show that MoIE does not hurt the accuracy of the original BB and eliminates shortcuts effectively.