A Fair Loss Function for Network Pruning

TL;DR

This paper introduces a performance weighted loss function to mitigate bias amplification during neural network pruning, ensuring fairness in resource-constrained deployment scenarios.

Contribution

It proposes a simple modified loss function that helps existing pruning methods reduce bias amplification, enhancing fairness in neural network pruning.

Findings

Effective in fairness-sensitive contexts

Reduces bias amplification during pruning

Compatible with existing pruning methods

Abstract

Model pruning can enable the deployment of neural networks in environments with resource constraints. While pruning may have a small effect on the overall performance of the model, it can exacerbate existing biases into the model such that subsets of samples see significantly degraded performance. In this paper, we introduce the performance weighted loss function, a simple modified cross-entropy loss function that can be used to limit the introduction of biases during pruning. Experiments using the CelebA, Fitzpatrick17k and CIFAR-10 datasets demonstrate that the proposed method is a simple and effective tool that can enable existing pruning methods to be used in fairness sensitive contexts. Code used to produce all experiments contained in this paper can be found at https://github.com/robbiemeyer/pw_loss_pruning.