CosFairNet:A Parameter-Space based Approach for Bias Free Learning

TL;DR

This work introduces a novel approach to address bias directly in the model's parameter space, preventing its propagation across layers, and shows enhanced classification accuracy and debiasing effectiveness across various synthetic and real-world datasets of different sizes.

Abstract

Deep neural networks trained on biased data often inadvertently learn unintended inference rules, particularly when labels are strongly correlated with biased features. Existing bias mitigation methods typically involve either a) predefining bias types and enforcing them as prior knowledge or b) reweighting training samples to emphasize bias-conflicting samples over bias-aligned samples. However, both strategies address bias indirectly in the feature or sample space, with no control over learned weights, making it difficult to control the bias propagation across different layers. Based on this observation, we introduce a novel approach to address bias directly in the model's parameter space, preventing its propagation across layers. Our method involves training two models: a bias model for biased features and a debias model for unbiased details, guided by the bias model. We enforce dissimilarity in the debias model's later layers and similarity in its initial layers with the bias model, ensuring it learns unbiased low-level features without adopting biased high-level abstractions. By incorporating this explicit constraint during training, our approach shows enhanced classification accuracy and debiasing effectiveness across various synthetic and real-world datasets of different sizes. Moreover, the proposed method demonstrates robustness across different bias types and percentages of biased samples in the training data. The code is available at: https://visdomlab.github.io/CosFairNet/

Figures (3)

• Figure 1: Gradient activation map of biased and debiased models in lower and higher layers for the BFFHQ dataset. (a) and (b) illustrate gender predictions of the biased and debiased models at the initial layer. It is evident that the debiased model closely corresponds to the biased model. In contrast, (c) and (d) show no correspondence as the learned weights differ at the higher layer. (Best view in color).
• Figure 2: The architecture of CosFairNet depicts the debiasing mechanism where $F_d$ and $F_b$ are the debias and bias models, respectively. W represents the weighting of samples. Cosine similarity is employed to align (initial layers) or de-align (later layers) $F_d$ and $F_b$ model layers to ensure differentiated learning of biased and unbiased representations. In figure, $CE(.)$ stands for cross-entropy loss, $GCE(.)$ for generalized cross-entropy loss and $W$ for relative difficulty score.
• Figure 3: t-SNE feature visualization comparing the vanilla model (without debiasing) and the proposed debias model on right side. Red and Green points represent samples from two different classes of the BFFHQ dataset. After debiasing with the proposed model, noticeable enhancements in class discrimination are observed, resulting in better separation of classes and hence, better classification accuracy. (Best viewed in color).