InfoCAM: An information-weighted class activation mapping for explaining visual neural networks

With the rapid advancement of deep learning technologies, visual neural networks have made significant progress across various benchmarks. However, these networks heavily rely on nonlinear functions and hyperparameter tuning techniques, which leads to black-box behavior during forward inference. To enhance the transparency of decision-making, the information flow in visual neural networks is revisited, and a saliency visual explanation framework called Information-weighted Class Activation Mapping (InfoCAM) is proposed. The InfoCAM regards deep visual neural networks as complex information processing systems and constructs a Dual-Stream Information Bottleneck (DSIB) module. DSIB adopts a divide-and-conquer strategy by introducing two latent variables and employs variational inference to explicitly decompose intermediate feature activations into a discriminative feature stream and a task-irrelevant noise stream. Through maximizing the mutual information between the discriminative stream and the output prediction, the weighting coefficients are faithfully assigned to feature activation maps, thereby generating reliable visual explanations. InfoCAM is not affected by the shattered gradient problem and can be seamlessly integrated into various visual network architectures and computer vision tasks. Furthermore, the average drop and average increase metrics used for faithfulness evaluation are improved by explicitly clarifying their theoretical bounds and mitigating their sensitivity to fluctuations in network output scores. Experimental results show that InfoCAM can achieve faithful explanation of visual neural networks, demonstrating superior performance in perturbation testing and energy-based pointing game evaluations.