# Overcoming Simplicity Bias in Deep Networks using a Feature Sieve

**Authors:** Rishabh Tiwari, Pradeep Shenoy  
**arXiv:** [2301.13293](https://arxiv.org/abs/2301.13293)  
**Date:** 2023-01-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2301.13293v3/

## Key points

- Deep neural networks (DNNs) have a simplicity bias, relying on weakly predictive features over stronger, more complex ones.
- Limited training data and spurious feature-label correlations compound this issue, leading to biased and incorrect predictions.
- The authors propose a direct interventional method called the feature sieve (SIFER) to address this problem in DNNs.
- SIFER identifies and suppresses easily computable spurious features in lower layers of the network, allowing higher levels to extract richer and more meaningful representations.
- SIFER outperforms many baselines that explicitly incorporate prior knowledge of spurious attributes or features.
- Using feature importance visualizations, the authors demonstrate that SIFER correctly identifies important visual features while suppressing irrelevant but spuriously label-correlated background features.
- SIFER encourages further research into designing computational barriers for neural networks by automating the extraction and combination of diverse features ordered by complexity and predictive power towards debiasing machine learning models.

## Summary

The phenomenon of simplicity bias in deep neural networks (DNNs) is a concerning tendency for these models to rely on simple, weakly predictive features at the expense of stronger, more complex ones. This issue is further compounded by limited training data and spurious feature-label correlations, leading to biased and incorrect predictions. In response, the authors propose a direct interventional method called the feature sieve (SIFER) to address this problem in DNNs. SIFER automatically identifies and suppresses easily computable spurious features in lower layers of the network, allowing higher levels to extract richer and more meaningful representations. The authors provide concrete evidence of differential suppression and enhancement of relevant features on both controlled datasets and real-world images. They report substantial gains on many real-world debiasing benchmarks such as an 11.4% relative gain on Imagenet-A and 3.2% on BAR. Importantly, SIFER does not require prior knowledge of spurious attributes or features and outperforms many baselines that explicitly incorporate such information. Furthermore, using feature importance visualizations, the authors demonstrate that SIFER correctly identifies important visual features while suppressing irrelevant but spuriously label-correlated background features. This underscores the relevance of SIFER to real-world feature understanding. In conclusion, this work with SIFER encourages further research into designing computational barriers for neural networks by automating the extraction and combination of diverse features ordered by complexity and predictive power towards debiasing machine learning models. The related work section highlights studies investigating the correlation and impact of shortcuts taken by neural networks when learning from data using stochastic gradient descent (SGD). Overall, this paper opens up exciting new research directions in automated adversarial feature extraction and representation learning for deep networks which could lead to improved accuracy in machine learning models through better understanding of their underlying structure.