[Paper Review] A Faster Approach to Spiking Deep Convolutional Neural Networks

Shahriar Rezghi Shirsavar, Mohammad‐Reza A. Dehaqani|arXiv (Cornell University)|Oct 31, 2022

Advanced Memory and Neural Computing4 citations

TL;DR

This paper proposes a faster, more accurate spiking deep convolutional neural network (SNN) by integrating single-training iteration, PCA-based dimensionality reduction, weight quantization to binary weights, timed outputs for classification, and improved preprocessing for RGB images. The method achieves 99.421% ± 0.033 accuracy on MNIST and outperforms ResNet on ETH-80, demonstrating significant speed and accuracy gains while maintaining biological plausibility.

ABSTRACT

Spiking neural networks (SNNs) have closer dynamics to the brain than current deep neural networks. Their low power consumption and sample efficiency make these networks interesting. Recently, several deep convolutional spiking neural networks have been proposed. These networks aim to increase biological plausibility while creating powerful tools to be applied to machine learning tasks. Here, we suggest a network structure based on previous work to improve network runtime and accuracy. Improvements to the network include reducing training iterations to only once, effectively using principal component analysis (PCA) dimension reduction, weight quantization, timed outputs for classification, and better hyperparameter tuning. Furthermore, the preprocessing step is changed to allow the processing of colored images instead of only black and white to improve accuracy. The proposed structure fractionalizes runtime and introduces an efficient approach to deep convolutional SNNs.

Motivation & Objective

• To close the accuracy and speed gap between spiking neural networks (SNNs) and deep neural networks (DNNs) by enhancing runtime and performance.
• To improve biological plausibility while maintaining high accuracy in SNNs through biologically inspired learning rules and coding schemes.
• To enable efficient deployment on neuromorphic hardware via binary weights and spike-based computation.
• To extend SNNs to colored image processing, improving accuracy beyond traditional grayscale-only approaches.
• To reduce training complexity by limiting training to a single iteration, enhancing computational efficiency.

Proposed method

• The network uses leaky integrate-and-fire (LIF) neurons with rank order coding for temporal encoding of input signals.
• Spike-timing-dependent plasticity (STDP) is applied for unsupervised, biologically plausible learning in convolutional layers.
• Principal component analysis (PCA) is used for dimensionality reduction to accelerate training and reduce computational load.
• Synaptic weights are quantized to binary values (±1), enabling efficient memory usage and fast binary multiplication.
• Timed output classification is employed, where the neuron with the highest spike count or earliest firing time determines the class.
• Input preprocessing is enhanced to support RGB images by converting to HSV color space, improving feature representation and accuracy.

Experimental results

Research questions

• RQ1 Can a single-training-iteration SNN achieve high accuracy comparable to state-of-the-art DNNs while maintaining biological plausibility?
• RQ2 How does PCA-based dimensionality reduction affect the runtime and accuracy of deep SNNs?
• RQ3 To what extent does RGB image processing via HSV transformation improve SNN performance on benchmark datasets?
• RQ4 Can binary weights and spike-based computation enable faster inference and lower energy consumption on neuromorphic hardware?
• RQ5 How does the proposed SNN compare to ResNet in accuracy and runtime on the ETH-80 dataset?

Key findings

• The proposed SNN achieves 99.421% ± 0.033 accuracy on the MNIST dataset using a single training iteration and SVM with RBF kernel.
• On the ETH-80 dataset, the HSV-preprocessed SNN achieves 90.166% accuracy, significantly outperforming the baseline ResNet-50 (86.585%).
• The use of ZCA whitening instead of LoG filtering reduces total runtime by 7.447–11.886 ms per sample while maintaining comparable accuracy.
• The network shows a significant accuracy improvement of 6.402% (from 72.378% to 78.780%) when using RGB (HSV) images over grayscale images on ETH-80.
• The 95% confidence interval for the difference in accuracy between LoG and ZCA whitening is [-0.841, 2.108], indicating no significant drop in accuracy with ZCA.
• The network’s performance is comparable to fine-tuned ResNet-50, with no significant difference in accuracy or runtime (95% CI: [-1.83, 5.874] and [-1.665, 1.665], respectively).