突破类脑模型性能瓶颈：校正频率偏置实现性能与能效双突破｜NeurIPS 2025

Core Insights - The article discusses the limitations of Spiking Neural Networks (SNNs) and introduces a new architecture called Max-Former that addresses these limitations by enhancing high-frequency information processing [5][24]. Group 1: Performance Limitations of SNNs - SNNs have been traditionally viewed as inferior to Artificial Neural Networks (ANNs) due to their binary pulse transmission, which was believed to cause significant information loss [5][6]. - The research indicates that the real issue lies in the frequency bias of SNNs, where spiking neurons act as low-pass filters, suppressing high-frequency components and favoring low-frequency information [4][8][19]. - This frequency imbalance leads to a degradation in the feature representation capabilities of SNNs, limiting their performance [10][23]. Group 2: Introduction of Max-Former - The Max-Former architecture is designed to counteract the inherent low-frequency preference of SNNs by incorporating two lightweight "frequency-enhancing lenses" [24][28]. - The architecture includes an additional Max-Pool operation in the Patch Embedding stage to actively inject high-frequency signals at the input source [28]. - It also replaces early-stage self-attention with deep convolution (DWC), which retains local high-frequency details while being computationally efficient [28]. Group 3: Performance Metrics and Results - Max-Former achieved a Top-1 accuracy of 82.39% on ImageNet with fewer parameters compared to Spikformer, demonstrating a significant performance improvement [27]. - The architecture also reduced energy consumption by over 30% while achieving performance breakthroughs [30]. - The findings suggest that optimizing SNNs with high-pass operators can lead to improvements in both performance and energy efficiency [31]. Group 4: Broader Implications - The insights gained from the Max-Former architecture are applicable beyond Transformer models, as demonstrated by the Max-ResNet architecture, which also benefited from the addition of high-frequency operations [33]. - The research provides a new perspective on the performance bottlenecks of SNNs, suggesting that their optimization should not merely mimic successful designs from ANNs [35].