Conversion of Continuous-Valued Deep Networks to Efficient Event-Driven Networks for Image Classification.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Spiking neural networks (SNNs) offer potential computational efficiency due to sparse activation and event-driven processing.
• Prior work enabled conversion of simple deep Convolutional Neural Networks (CNNs) to spiking equivalents.
• Common CNN operations like max-pooling, softmax, batch-normalization, and Inception-modules were not previously supported for SNN conversion.

Purpose of the Study:

• To develop spiking equivalents for common CNN operations, enabling the conversion of nearly arbitrary CNN architectures into SNNs.
• To demonstrate the effectiveness of these new conversion methods on popular CNN architectures.
• To highlight the trade-offs between classification error and operational efficiency in SNNs.

Main Methods:

• Introduced spiking equivalents for max-pooling, softmax, batch-normalization, and Inception-modules.
• Converted popular CNN architectures, including VGG-16 and Inception-v3, into SNNs.
• Evaluated SNN performance on MNIST, CIFAR-10, and ImageNet datasets.

Main Results:

• Achieved state-of-the-art results on MNIST, CIFAR-10, and ImageNet datasets with converted SNNs.
• Demonstrated that SNNs can achieve over 2x reduction in operations compared to original CNNs (e.g., LeNet, BinaryNet) with a slight increase in error rate.
• Showcased the potential for significant operational savings in SNNs.

Conclusions:

• The developed methods allow for the conversion of a wider range of CNN architectures into accurate and efficient SNNs.
• SNNs present a viable approach for reducing computational load, particularly for power-efficient neuromorphic hardware in embedded systems.
• The findings underscore the potential of SNNs for energy-efficient AI applications.