Bias in Pruned Vision Models: In-Depth Analysis and Countermeasures

Yet, several recent poorly on "unusual" data, which can frequently coincide works have raised the issue that pruning may induce or exacerbate with marginalized groups. Given the recent popularity of bias in the output of the compressed model. Despite compression methods in deployment settings [13,18,19,27] existing evidence for this phenomenon, the relationship and the fact that, for massive models, compression is often between neural network pruning and induced bias is necessary to enable model deployment, these findings raise not well-understood. In this work, we systematically investigate the question of whether the bias due to compression can be and characterize this phenomenon in Convolutional exactly characterized, and in particular whether bias is an Neural Networks for computer vision. First, we show that it inherent side-effect of the model compression process. is in fact possible to obtain highly-sparse models, e.g. with In this paper, we perform an in-depth analysis of bias in less than 10% remaining weights, which do not decrease in compressed vision models, providing new insights on this accuracy nor substantially increase in bias when compared phenomenon, as well as a set of practical, effective criteria to dense models. At the same time, we also find that, at for identifying samples susceptible to biased predictions, higher sparsities, pruned models exhibit higher uncertainty which can be used to significantly attenuate bias. in their outputs, as well as increased correlations, which Our work starts from a common setting to study bias we directly link to increased bias. We propose easy-to-use and bias mitigation [28, 29, 40, 50]: we study properties of criteria which, based only on the uncompressed model, establish sparse residual convolutional neural networks [25], in particular whether bias will increase with pruning, and identify ResNet18, applied for classification on the CelebA the samples most susceptible to biased predictions postcompression.