Evaluating Robustness of Pre-Trained Deep Neural Networks against Spurious Correlations

Abstract

This paper addresses the challenge of sub-optimal performance in deep neural networks trained for image classification under non-matching distribution scenarios. Spurious correlations, patterns in training data irrelevant to objects, can lead to accuracy loss during testing. We assess the robustness of pre-trained models to spurious correlations by subjecting them to datasets containing such correlations. We compare the effectiveness of two training methods; fine-tuning and backbone freezing. Additionally, we explore the impact of robust training on our model collection by applying the DFR method to both frozen and fine-tuned model backbones.

Alireza Hoseinpour

Masters Student in Computer Science

My research interests include AI for Software Engineering, Software Engineering for AI, Software Maintenance and Evolution, Software Testing, Software Analytics, and Empirical Software Engineering.