# ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness

**Authors:** Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, Wieland Brendel  
**arXiv:** [1811.12231](https://arxiv.org/abs/1811.12231)  
**Date:** 2018-11-29  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/1811.12231v1/

## Key points

- Study titled "ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness"
- Researchers evaluated CNNs and human observers on images with a texture-shape cue conflict
- ImageNet-trained CNNs exhibit a strong bias towards recognizing textures over shapes
- Training the network on a stylized version of ImageNet can shift its representation from texture-based to shape-based
- Shape-based representation leads to improved object detection and enhanced robustness against image distortions
- Nine experiments totaling 48,560 psychophysical trials across 97 observers in a well-controlled lab setting
- Advantages of adopting a shape-based representation in CNNs for accurate and robust object recognition tasks
- Study currently under review at ICLR 2019 with favorable scores (8, 8, 7)
- Implications for improving machine learning algorithms

## Summary

In their study titled "ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness," authors Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel delve into the classification strategies of Convolutional Neural Networks (CNNs) in recognizing objects. The researchers evaluated CNNs and human observers on images with a texture-shape cue conflict to test conflicting hypotheses about the role of textures and shapes in object recognition. Their findings reveal that ImageNet-trained CNNs exhibit a strong bias towards recognizing textures over shapes, highlighting fundamentally different classification strategies employed by machines compared to humans. However, the authors demonstrate that training the network on a stylized version of ImageNet can shift its representation from texture-based to shape-based. This not only aligns more closely with human performance but also leads to improved object detection and enhanced robustness against image distortions. Through nine experiments totaling 48,560 psychophysical trials across 97 observers in a well-controlled lab setting, the study showcases the advantages of adopting a shape-based representation in CNNs for accurate and robust object recognition tasks. Currently under review at ICLR 2019 with favorable scores (8, 8, 7), this study sheds light on the mechanisms underlying object recognition in neural networks and has implications for improving machine learning algorithms.