Lightweight spiking transformer towards neurodynamic integration framework

Spiking Neural Networks (SNNs) have gained significant attention in recent years due to their high energy efficiency and event-driven characteristics. Meanwhile, Transformer architectures have demonstrated strong capabilities in modeling global dependencies. However, effectively integrating Transformers into SNNs remains challenging, primarily due to the high computational cost and the spiking computing paradigm. Existing spiking Transformer models still rely on conventional self-attention mechanisms, resulting in low computational efficiency on resource constrained hardware. To address these challenges, we proposes a unified neuro-dynamic spiking Transformer model, termed Lightweight Spiking Transformer (LST). Specifically, LST comprises three key components. First, a Spiking Visual Tokenizer (SVT) replaces max pooling operations with downsampling, thereby reducing information loss. Second, a Spiking Lambda Attention (SLA) mechanism reformulates attention computation via global context aggregation, enabling effective modeling of long range dependencies while avoiding the explicit construction of attention matrices. Finally, a Dual-Threshold Adaptive LIF (DTA-LIF) neuron incorporates adaptive threshold dynamics and a dual-branch firing structure to enhance representational flexibility of input signals. By jointly optimizing input representation, attention modeling, and neuron dynamics, the proposed model achieves an efficient and hardware friendly design. Extensive experiments demonstrate that the proposed model achieves 61.15%,96.11%, 79.24%, and 98.4% top-1 accuracy on Tiny-Imagenet, CIFAR10, CIFAR100, and DVS128 Gesture datasets, respectively, while reducing the parameter count by 20.8%, 3.4%, and 3.8% compared with state-of-the-art Spiking Transformer models. These results indicate that the proposed framework effectively balances performance and energy efficiency, providing a promising solution for high-performance, low-power visual processing using SNNs.