Spatial Information Bottleneck for Interpretable Visual Recognition

The introduction of the Spatial Information Bottleneck (S-IB) framework marks a significant advancement in the field of deep learning, particularly in visual recognition. Traditional deep neural networks often struggle with interpretability due to their tendency to conflate discriminative foreground features with spurious background correlations. S-IB addresses this challenge by optimizing the Vector-Jacobian Products (VJP) during backpropagation, which allows for a clearer separation of relevant information. By maximizing mutual information between foreground VJP and inputs while minimizing it in background regions, S-IB encourages networks to focus on class-relevant spatial areas. Experiments conducted on five benchmarks have shown improvements in model performance and visualization quality across six explanation methods, underscoring the potential of S-IB to enhance the interpretability and robustness of neural networks.