Bias Alleviation Through Network Pruning for Sparse and Debiased Models.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Neural network pruning effectively reduces model size with minimal performance loss.
• Pruning can exacerbate existing biases, negatively impacting performance for underrepresented groups.
• Existing debiasing methods often require sensitive group information.

Purpose of the Study:

• To introduce a novel method for identifying and mitigating bias amplification caused by neural network pruning.
• To develop a debiasing algorithm that does not require explicit group annotations.
• To achieve both model sparsity and bias reduction simultaneously.

Main Methods:

• Introduced Accumulated Confidence (AC) as a proxy to identify bias-aligned and bias-conflicting samples.
• Proposed the DEbiasing Network through Pruning (DENP) algorithm, leveraging AC for bias mitigation.
• Evaluated DENP's performance across various sparsity levels and benchmark datasets.

Main Results:

• Accumulated Confidence (AC) effectively identifies bias without relying on group annotations.
• The DEbiasing Network through Pruning (DENP) algorithm successfully mitigated pruning-induced bias.
• DENP achieved superior debiasing performance compared to state-of-the-art methods, even at high sparsity levels.
• Resulting models were both sparse and debiased.

Conclusions:

• Pruning's bias-exacerbating effects can be effectively addressed using the proposed AC metric and DENP algorithm.
• DENP offers a robust solution for creating fair and efficient neural networks.
• The method demonstrates significant potential for improving model fairness in real-world applications.