You Never Know: Quantization Induces Inconsistent Biases in Vision-Language Foundation Models
This addresses fairness concerns in deploying compressed AI models, revealing inconsistent bias effects that challenge prior assumptions.
The study investigated how quantization affects social bias in vision-language foundation models, finding that while individual models show bias, there is no consistent change in bias magnitude or direction across compressed models due to quantization.
We study the impact of a standard practice in compressing foundation vision-language models - quantization - on the models' ability to produce socially-fair outputs. In contrast to prior findings with unimodal models that compression consistently amplifies social biases, our extensive evaluation of four quantization settings across three datasets and three CLIP variants yields a surprising result: while individual models demonstrate bias, we find no consistent change in bias magnitude or direction across a population of compressed models due to quantization.