Downsized and Compromised?: Assessing the Faithfulness of Model Compression

High accuracy alone cannot guarantee that compressed models behave the same way as their original versions. The paper found that models maintaining similar overall accuracy scores still exhibited significant changes in their decision-making patterns after compression, which is particularly concerning in high-stakes domains like healthcare, finance, and criminal justice.

Compression techniques can introduce hidden biases that standard metrics miss. Using statistical tests (chi-squared tests), the researchers detected meaningful shifts in how compressed models make predictions across different demographic subgroups—changes that traditional fairness metrics might overlook, potentially disadvantaging specific populations.

Model agreement between original and compressed versions is a critical but overlooked metric. The paper introduces techniques to directly measure whether compressed and original models make the same predictions on the same inputs, revealing faithfulness problems that accuracy measurements alone cannot capture.

Compression trade-offs require more rigorous evaluation in high-stakes applications. While model compression reduces computational costs and size, the paper demonstrates that current evaluation practices are insufficient for domains where consistent, predictable model behavior is essential for responsible deployment.

Demographic subgroup analysis is essential when evaluating compressed models. The research shows that compression effects are not uniform across populations—some demographic groups may experience larger shifts in model behavior than others, requiring disaggregated testing beyond aggregate performance metrics.