Problem

Spiking Neural Networks (SNNs) offer biological plausibility and energy efficiency, particularly on neuromorphic hardware, but struggle to achieve state-of-the-art performance on vision tasks compared to traditional Artificial Neural Networks (ANNs). Training SNNs is challenging due to the discrete, non-differentiable nature of spikes, which prevents direct application of standard backpropagation. While surrogate gradient methods have emerged, there's still a need for effective spike coding strategies to fully leverage SNN advantages for improved performance in complex vision problems.

Method

This study introduces a novel hybrid temporal-8-bit spike coding method designed for surrogate gradient training of SNNs. The approach integrates bit-plane decomposition of input images with time-to-first-spike (TTFS) temporal coding. For an input image, pixel intensities are first decomposed into their individual bit planes (e.g., 8-bit image yields 8 bit planes). Each bit plane is a binary tensor. Concurrently, the original image intensities are encoded using TTFS, where spike times are inversely proportional to the input magnitude (larger intensity means earlier spike). The final spike-chain representation is formed by concatenating the TTFS-encoded spike times with the extracted bit-plane tensors along the temporal dimension. The SNNs are trained using hard-reset integrate-and-fire (IF) neurons with an arctangent surrogate gradient function and backpropagation through time (BPTT), optimized with Adam and cross-entropy loss.

Results

Extensive experiments demonstrated that the proposed hybrid temporal-8-bit coding method yields competitive and often superior performance across various computer vision benchmarks. On grayscale datasets like MNIST, KMNIST, and Fashion-MNIST, it achieved the highest accuracy on MNIST (95.65%) and an average accuracy improvement of 1.09% over the TTFS baseline, outperforming a prior hybrid rate-8-bit approach. When evaluated against state-of-the-art SNN architectures and optimization algorithms (Adam and SGD) on CIFAR10, the method consistently achieved higher accuracy than baselines, with performance gains ranging from 0.15% to 17.6%. For color image classification on datasets such as CIFAR10, CIFAR100, EuroSAT, Caltech101, Caltech256, Imagenette, and Food101, the proposed method consistently surpassed baselines, showing a more significant benefit with gains from 0.62% (EuroSAT) to 19.6% (Food101), and outperforming the prior hybrid rate-8-bit approach by an average of 0.23%.

Implications

The development of this hybrid temporal-8-bit spike coding method represents a significant step towards improving the performance of Spiking Neural Networks. By effectively integrating bit-plane information with temporal coding, the method enhances SNN capabilities in various computer vision tasks. Its demonstrated generalization across different SNN architectures and optimization algorithms suggests its robustness and potential for broader adoption. These findings pave the way for more efficient, energy-saving, and effective SNN models, particularly relevant for neuromorphic computing and real-time AI applications, potentially bridging the performance gap with traditional ANNs.