Rectifying Shortcut Learning through Cellular Differentiation in Deep Learning Neurons


Hongjing Niu (University of Science and Technology of China), Hanting Li (University of Science and Technology of China), Guoping Wu (University of Science and Technology of China), Bin Li (University of Science and Technology of China), Feng Zhao (University of Science and Technology of China)
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)Supplementary

Abstract

Deep learning models have exhibited tendencies to rely on shortcuts, such as identifying sheep based solely on the presence of grassland. This reliance often overshadows the true potential of deep learning, significantly impacting their performance and robustness in diverse scenarios. Particularly, shortcuts embedded in training data may prove detrimental in testing scenarios, acting as sources of interference rather than assistance. A notable challenge arises in biased datasets, where the learning of the intended attribute is hindered by spurious correlations with extraneous attributes. This paper introduces an innovative method that harnesses positive feedback to foster internal differentiation within the model, thereby diminishing feature entanglement and enhancing the learning of distinct, non-shortcut features. Our approach employs model parameter masks to segment the model into locally specialized subcomponents. This segmentation facilitates the amplification of differences between these components through interaction, allowing for the distribution of the model's overall functionality across each subcomponent. In essence, our method enables the independent learning of diverse features, circumventing the model's reliance on shortcut features. Comprehensive experiments conducted on a variety of datasets from multiple perspectives have demonstrated the efficacy of our proposed method. It significantly enriches feature diversity and improves model performance in complex scenarios.

Citation

@inproceedings{Niu_2024_BMVC,
author    = {Hongjing Niu and Hanting Li and Guoping Wu and Bin Li and Feng Zhao},
title     = {Rectifying Shortcut Learning through Cellular Differentiation in Deep Learning Neurons},
booktitle = {35th British Machine Vision Conference 2024, {BMVC} 2024, Glasgow, UK, November 25-28, 2024},
publisher = {BMVA},
year      = {2024},
url       = {https://papers.bmvc2024.org/0736.pdf}
}


Copyright © 2024 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection