HSA²M: Hierarchical Spike Aggregation Activation Map for Visual Explanations From Spiking Neural Networks

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Spiking neural networks (SNNs) present challenges in visual explanation due to spike sparsity and temporal discontinuity.
• Existing methods struggle with generating coherent saliency maps, limiting SNN interpretability.
• Neurobiological principles offer potential solutions for enhancing SNN explainability.

Purpose of the Study:

• To develop a novel method, Hierarchical Spike Aggregation Activation Map (HSA²M), for generating fine-grained visual explanations in SNNs.
• To address the limitations of current SNN interpretability techniques.
• To leverage hierarchical feature integration and interspike interval saliency from neuroscience.

Main Methods:

• HSA²M employs multilayer spike aggregation and adaptive fusion, inspired by the ventral visual pathway and short interspike intervals.
• Key modules include a Spike Activation Map Generator (SAMG) for layer-wise saliency, Fisher-Weighted Fusion (FWF) for adaptive map integration using Fisher Information (FI), and a Metric-Aware Hyperparameter Optimizer (MA-HPO).
• The method maintains event-driven efficiency inherent to spike-based processing.

Main Results:

• HSA²M significantly improved interpretability, demonstrated by a 4.81% increase in ADCC.
• Faithfulness was enhanced, with Spearman's rank correlation coefficient (ρ) improving by 10.085%.
• Adversarial robustness showed marked improvement, with normalized L1 distance decreasing by 86.32%.

Conclusions:

• HSA²M provides precise and high-fidelity visual explanations for SNNs, outperforming state-of-the-art methods.
• The method successfully integrates neurobiological principles into SNN interpretability.
• HSA²M offers a promising approach for advancing the explainability and trustworthiness of SNNs across various benchmarks.