[Paper Review] Direct training based spiking convolutional neural networks for object recognition.

Shibo Zhou, Ying Chen+2 more|arXiv (Cornell University)|Sep 24, 2019

Advanced Memory and Neural Computing4 citations

TL;DR

This paper proposes a direct training-based spiking convolutional neural network (SCNN) using temporal coding to achieve high object recognition accuracy with minimal trainable parameters. It attains 99.13% accuracy on MNIST and 80.49% on CIFAR-10—state-of-the-art for direct-trainable SNNs with temporal coding—while significantly reducing parameter count compared to conversion-based methods.

ABSTRACT

Direct training based spiking neural networks (SNNs) have been paid a lot of attention recently because of its high energy efficiency on emerging neuromorphic hardware. However, due to the non-differentiability of the spiking activity, most of the related SNNs still cannot achieve high object recognition accuracy for the complicated dataset, such as CIFAR-10. Even though some of them can reach the accuracy of 90%, the energy consumption in those networks is very high. Considering this, we propose a direct supervised learning based spiking convolutional neural networks (SCNNs) using temporal coding scheme in this study, aiming to exploit minimum trainable parameters to recognize the object in the image with high accuracy. The MNIST and CIFAR-10 datasets are used to evaluate the performance of the proposed networks. For the MNIST dataset, the proposed networks with noise input are able to reach the high recognition accuracy (99.13%) as the other state-of-art models but use the much less trainable parameters than them. For CIFAR-10 dataset, the proposed networks with data augmentation step can reach the recognition accuracy of 80.49%., which is the state-of-art high accuracy in the field of direct training based SNNs using temporal coding manner. In addition, the number of trainable parameters used in such networks is much less than that in the conversion based SCNNs reported in the literature.

Motivation & Objective

• Address the challenge of low accuracy in direct training-based spiking neural networks (SNNs) on complex datasets like CIFAR-10.
• Reduce the number of trainable parameters in SNNs to improve energy efficiency and scalability.
• Achieve high object recognition accuracy using temporal coding in a direct supervised learning framework.
• Outperform existing direct-training SNNs in accuracy while minimizing parameter count, especially compared to conversion-based approaches.
• Demonstrate the feasibility of high-accuracy SNNs on standard benchmarks using minimal trainable parameters.

Proposed method

• Employ a direct supervised learning approach to train spiking convolutional neural networks (SCNNs) without conversion from artificial neural networks.
• Use a temporal coding scheme to represent input data, where information is encoded in the timing of spikes.
• Introduce noise input during training to enhance robustness and improve generalization on MNIST.
• Apply data augmentation techniques to improve performance on the more complex CIFAR-10 dataset.
• Optimize the network with backpropagation through time (BPTT) or similar differentiable surrogate gradients to handle non-differentiable spiking neurons.
• Minimize the number of trainable parameters by designing a compact network architecture tailored for spiking dynamics.

Experimental results

Research questions

• RQ1 Can a direct training-based SNN with temporal coding achieve high accuracy on MNIST and CIFAR-10 without conversion from ANNs?
• RQ2 How does the inclusion of noise input during training affect the generalization and accuracy of the proposed SNN?
• RQ3 To what extent can the number of trainable parameters be reduced while maintaining high recognition performance?
• RQ4 How does the proposed direct-training SNN compare to state-of-the-art conversion-based SNNs in terms of accuracy and parameter efficiency?
• RQ5 Can data augmentation improve the performance of direct-trainable SNNs on complex datasets like CIFAR-10?

Key findings

• The proposed SNN achieves 99.13% recognition accuracy on the MNIST dataset, matching state-of-the-art models while using significantly fewer trainable parameters.
• On the CIFAR-10 dataset, the model reaches 80.49% accuracy—representing the state-of-the-art performance among direct-training SNNs using temporal coding.
• The number of trainable parameters in the proposed network is substantially lower than in conversion-based SCNNs reported in the literature.
• The use of noise input during training enhances model robustness and contributes to high accuracy on MNIST.
• Data augmentation is critical for achieving high performance on the more complex CIFAR-10 dataset with direct-trainable SNNs.
• The proposed method demonstrates that high accuracy in SNNs is achievable with minimal parameter counts through direct training and temporal coding.