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

We demonstrate that ‘stereotyping’, i.e., amplifying feature correlations, increases with model sparsity, thus leading to increased bias.

Download paper here