[论文解读] Direct Training for Spiking Neural Networks: Faster, Larger, Better

Yujie Wu, Lei Deng+3 more|arXiv (Cornell University)|Sep 16, 2018

Advanced Memory and Neural Computing被引用 106

一句话总结

本文提出一种直接训练深度脉冲神经网络（SNN）的框架，使用显式迭代的 LIF 模型、NeuNorm 归一化和优化的速率编码，实现大规模、训练更快的 SNN，在神经形态数据集和非脉冲数据集上达到有竞争力的准确率。

ABSTRACT

Spiking neural networks (SNNs) that enables energy efficient implementation on emerging neuromorphic hardware are gaining more attention. Yet now, SNNs have not shown competitive performance compared with artificial neural networks (ANNs), due to the lack of effective learning algorithms and efficient programming frameworks. We address this issue from two aspects: (1) We propose a neuron normalization technique to adjust the neural selectivity and develop a direct learning algorithm for deep SNNs. (2) Via narrowing the rate coding window and converting the leaky integrate-and-fire (LIF) model into an explicitly iterative version, we present a Pytorch-based implementation method towards the training of large-scale SNNs. In this way, we are able to train deep SNNs with tens of times speedup. As a result, we achieve significantly better accuracy than the reported works on neuromorphic datasets (N-MNIST and DVS-CIFAR10), and comparable accuracy as existing ANNs and pre-trained SNNs on non-spiking datasets (CIFAR10). {To our best knowledge, this is the first work that demonstrates direct training of deep SNNs with high performance on CIFAR10, and the efficient implementation provides a new way to explore the potential of SNNs.

研究动机与目标

• Motivate and address why direct training of deep SNNs remains challenging compared to ANNs.
• Propose algorithmic advances (NeuNorm, rate coding optimization) to improve training and convergence.
• Provide a PyTorch-compatible, explicitly iterative LIF formulation to enable large-scale SNNs.
• Demonstrate significant training acceleration and competitive accuracy on neuromorphic and non-spiking datasets.

提出的方法

• Convert the Leaky Integrate-and-Fire (LIF) model to an explicitly iterative form for compatibility with ML frameworks.
• Introduce NeuNorm to balance neural selectivity by normalizing neuronal activity across feature maps in a layer.
• Optimize rate coding for input encoding and output decoding to reduce required simulation length.
• Integrate NeuNorm and the modified LIF into the spatio-temporal backpropagation (STBP) framework for direct training.
• Provide pseudo-code and a PyTorch implementation (Algorithm 2) for end-to-end training of deep SNNs.

实验结果

研究问题

• RQ1 Can direct training of deep SNNs achieve competitive accuracy with ANN-based benchmarks on both neuromorphic and non-spiking datasets?
• RQ2 Does NeuNorm improve neural selectivity and training convergence for deep SNNs?
• RQ3 How much speedup and scalability can be achieved by using an explicitly iterative LIF model and PyTorch-based implementation?
• RQ4 What is the impact of optimized rate coding on training efficiency and accuracy across datasets like N-MNIST, DVS-CIFAR10, and CIFAR10?

主要发现

• Direct training of deep SNNs (up to 8 layers) yields superior accuracy on neuromorphic datasets compared with prior SNN work.
• NeuNorm improves training convergence and classification performance by balancing activity across feature maps.
• A PyTorch-implemented explicitly iterative LIF model provides tens of times speedup over Matlab-based implementations.
• The approach achieves best reported accuracy on N-MNIST (with NeuNorm: 99.53%) and DVS-CIFAR10 (60.5%), and competitive CIFAR10 results (90.53% with NeuNorm).
• The encoding/decoding scheme reduces the required simulation length (4–8 steps) while preserving performance.
• Network scaling shows accuracy improves with larger architectures, similar to findings in ANN literature.