Simon Says: Evaluating and Mitigating Bias in Pruned Neural Networks with Knowledge Distillation

Neural network pruning can inadvertently create bias: The paper identifies that when AI models are compressed through pruning (a technique that removes parts of neural networks to make them faster), this process can introduce or amplify bias—a problem that hasn't been well-studied until now.

Knowledge distillation offers a practical solution: The researchers demonstrate that knowledge distillation (a technique where a smaller model learns from a larger one) can effectively reduce bias in pruned models, even when training data is imbalanced, providing practitioners with a concrete mitigation strategy.

New metrics enable bias measurement in compressed models: The paper proposes two new measurement tools—Combined Error Variance (CEV) and Symmetric Distance Error (SDE)—that allow organizations to quantitatively assess whether their pruned AI systems are treating different groups fairly, addressing a gap in current evaluation methods.

Model similarity predicts where bias will occur: The research reveals that how similar a pruned model is to the original model correlates strongly with bias levels, offering an explainable way to understand why bias emerges during compression—crucial for building trustworthy AI systems.

Bias in compressed AI is harder to detect than human bias: The authors emphasize that algorithmic bias from model compression is "more abstract and unintuitive" than traditional discrimination, making it essential to develop specialized evaluation tools before deploying these optimized models in real-world applications.