Related Experiment Video
Updated: Jan 9, 2026
Closed-loop Neuro-robotic Experiments to Test Computational Properties of Neuronal Networks
Published on: March 2, 2015
Towards efficient and accurate spiking neural networks via adaptive bit allocation
Yao Xingting1, Hu Qinghao2, Zhou Fei3
1The Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences, Zhongguan Cun East Road No.95, Beijing, 100190, China; School of Future Technology, University of Chinese Academy of Sciences, Huaibei Zhuang No.380, Beijing, 101408, China.
This study introduces an adaptive bit allocation strategy for multi-bit spiking neural networks (SNNs), optimizing layer-wise resource use. This approach enhances SNNs' energy efficiency and accuracy by intelligently managing memory and computation demands.
Area of Science:
- Artificial Intelligence
- Computer Science
- Neuroscience
Background:
- Multi-bit Spiking Neural Networks (SNNs) offer energy-efficient AI but face escalating memory and computation demands with increased bit precision.
- Disproportionate performance gains arise from inefficient resource allocation in SNNs, where extra bits can be wasteful or detrimental.
Purpose of the Study:
- To develop an adaptive bit allocation strategy for direct-trained SNNs to achieve fine-grained, layer-wise optimization of memory and computation resources.
- To enhance the overall efficiency and accuracy of SNNs through intelligent resource management.
Main Methods:
- Parametrization of temporal lengths and bit widths for weights and spikes, making them learnable and controllable via gradients.
- Introduction of a refined spiking neuron capable of handling variable temporal lengths and enabling gradient derivation for temporal optimization.
- Theoretical formulation and resolution of the step-size mismatch problem in learnable bit widths using a step-size renewal mechanism to mitigate quantization errors.
Main Results:
- Demonstrated reduction in overall memory and computation costs across static (CIFAR, ImageNet) and dynamic (CIFAR-DVS, DVS-GESTURE) datasets.
- Achieved higher accuracy compared to baseline methods.
- SEWResNet-34 model on ImageNet showed a 2.69% accuracy gain and a 4.16x reduction in bit budgets.
Conclusions:
- The proposed adaptive bit allocation strategy effectively optimizes SNN resource utilization, leading to improved energy efficiency and accuracy.
- The novel components, including the refined spiking neuron and step-size renewal mechanism, successfully address challenges associated with variable bit widths and temporal lengths.
- This work provides a significant advancement in the practical deployment of high-performance, resource-efficient SNNs.
