Learning Not to Learn: Training Deep Neural Networks with Biased Data

TL;DR

This paper introduces a regularization method that uses mutual information minimization and adversarial training to prevent deep neural networks from learning biased data, improving generalization.

Contribution

It presents a novel regularization algorithm that unlearns bias in neural networks by adversarially minimizing mutual information between features and bias.

Findings

Effective removal of bias information from feature embeddings.

Improved test performance on biased datasets.

Quantitative and qualitative validation of bias unlearning.

Abstract

We propose a novel regularization algorithm to train deep neural networks, in which data at training time is severely biased. Since a neural network efficiently learns data distribution, a network is likely to learn the bias information to categorize input data. It leads to poor performance at test time, if the bias is, in fact, irrelevant to the categorization. In this paper, we formulate a regularization loss based on mutual information between feature embedding and bias. Based on the idea of minimizing this mutual information, we propose an iterative algorithm to unlearn the bias information. We employ an additional network to predict the bias distribution and train the network adversarially against the feature embedding network. At the end of learning, the bias prediction network is not able to predict the bias not because it is poorly trained, but because the feature embedding network successfully unlearns the bias information. We also demonstrate quantitative and qualitative experimental results which show that our algorithm effectively removes the bias information from feature embedding.