Brain-Like Object Recognition Neural Networks are more robustness to common corruptions

Anonymous

Keywords: Model Robustness, Neural Predictability, Representational Similarity

TL;DR: Brain-like deep learning models are more robustness to common corruptions

Abstract: Previous work (Schrimpf et al, 2018, Schrimpf et al, 2020) has shown that there exists a correlation between the performance of neural networks in object recognition tasks and its ability to match behavioral and neural recordings. We expanded on this work to ask the question: Does the behavioral and neural recordings are also correlated to the robustness of neural networks to common corruptions (e.g ImageNet-C). We selected several models from the leaderboard in Brain-Score, a platform that hosts neural and behavioral benchmarks for brain-model similarity, and tested their robustness to the corruption from ImageNet-C. We showed that higher brain-score is correlated with lower mean corruption error across models. Particularly, we show a correlation between the V4 and Behavioral datasets and the model's robustness to ImageNet-C. These finds suggest that explicitly modeling/matching data from V4 might be a good strategy for developing robust models to common corruptions.