Exploiting nonlinear dendritic adaptive computation in training deep Spiking Neural Networks.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence

Background:

• Spiking Neural Networks (SNNs) are inspired by brain information processing and are event-driven.
• Complex SNNs face challenges in managing spiking behavior, leading to information loss with dense or sparse spikes.
• Existing SNNs use linear summation, neglecting dendritic adaptive processing.

Purpose of the Study:

• To introduce novel modules that enhance information processing in SNNs.
• To address the limitations of linear summation in current SNNs.
• To improve the performance of SNNs on various datasets.

Main Methods:

• Introduced the Dendritic Spatial Gating Module (DSGM) to scale and translate inputs, reducing spike transformation loss.
• Implemented the Dendritic Temporal Adjust Module (DTAM) for assigning importance to inputs across time steps.
• Fused DSGM and DTAM to achieve balanced spike representation and integrate multi-step temporal information.

Main Results:

• Achieved state-of-the-art performance on static image datasets (CIFAR10, CIFAR100) and event datasets (DVS-CIFAR10, DVS-Gesture, N-Caltech101).
• Demonstrated competitive performance against state-of-the-art methods on the ImageNet dataset.
• Showcased enhanced neural network performance through balanced spike representation.

Conclusions:

• The proposed DSGM and DTAM modules significantly improve SNN performance.
• The novel approach effectively addresses information loss and temporal dependency challenges in SNNs.
• This work advances SNN capabilities for complex computational tasks.