Spatial Information Bottleneck for Interpretable Visual Recognition

Kaixiang Shu, Kai Meng, Junqin Luo

Deep neural networks typically learn spatially entangled representations that conflate discriminative foreground features with spurious background correlations, thereby undermining…