Breaking Global Self-Attention Bottlenecks in Transformer-based Spiking Neural Networks with Local Structure-Aware Self-Attention
Researchers have developed a novel Transformer-based Spiking Neural Network called LSFormer, designed to overcome limitations in existing models. LSFormer introduces Spiking Response Pooling (SPooling) and Local Structure-Aware Spiking Self-Attention (LS-SSA) to better preserve regional features and reduce computational redundancy. This new architecture utilizes a local dilated window mechanism to capture both fine-grained details and broader dependencies, achieving state-of-the-art results on datasets like Tiny-ImageNet and N-CALTECH101. AI
IMPACT Introduces a more efficient and accurate architecture for spiking neural networks, potentially enabling wider adoption in energy-constrained applications.