A structure-time parallel implementation of spike-based deep learning.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Deep Learning

Background:

• Recent advancements in deep spiking neural networks (SNNs) show promise for efficient AI.
• Existing spike-based deep learning algorithms, like broadcast alignment, can be computationally intensive.
• There is a need for accelerated methods to facilitate SNN research and application.

Purpose of the Study:

• To propose a structure-time parallel strategy to accelerate the broadcast alignment algorithm for deep SNNs.
• To develop a deep hierarchical model utilizing the parallel broadcast alignment for object recognition tasks.
• To enhance the efficiency and accuracy of spike-based deep learning models.

Main Methods:

• Implementation of a structure-time parallel strategy leveraging layered structures and one-time computation within a time window.
• Development of a deep hierarchical model incorporating the parallel broadcast alignment for object recognition.
• Evaluation on the MNIST dataset for speedup and the ETH-80 dataset for object recognition accuracy.

Main Results:

• The parallel broadcast alignment achieved a substantial 137x speedup compared to the original implementation on the MNIST dataset.
• The proposed object recognition model demonstrated higher accuracy than state-of-the-art spiking deep convolutional neural networks on the ETH-80 dataset.
• The strategy effectively accelerates deep SNN simulations and improves performance in real-world applications.

Conclusions:

• The proposed structure-time parallel strategy significantly enhances the speed of spike-based deep learning algorithms.
• The deep hierarchical model based on this strategy achieves competitive accuracy in object recognition.
• This work facilitates both the study of neural dynamics in deep SNNs and their practical applications.