Characterising Bias in Compressed Models

Compression techniques hide fairness problems: While pruning and quantization maintain overall accuracy metrics, they disproportionately harm performance on underrepresented groups and features—meaning a model can appear accurate on average while failing systematically for minority populations.

A small set of "problem examples" drives most errors: The researchers identified Compression Identified Exemplars (CIE)—a small subset of examples where compressed models fail at much higher rates than uncompressed ones—suggesting that focusing audits on these cases could efficiently surface bias issues.

Existing biases get amplified during compression: The paper demonstrates that compression doesn't just maintain existing algorithmic bias; it actively amplifies it, particularly affecting underrepresented features that often relate to fairness concerns like demographic groups.

Human review of compressed models is essential: The authors propose using CIE as a practical auditing tool where domain experts can manually inspect the most problematic examples, making bias detection more tractable than reviewing entire datasets.

Standard accuracy metrics are insufficient for responsible AI: The finding that top-1 and top-5 accuracy can remain "negligible" while fairness degrades significantly shows that companies deploying compressed models need disaggregated performance metrics across demographic groups, not just overall accuracy scores.