Supervised Learning Algorithm Based on Spike Train Inner Product for Deep Spiking Neural Networks.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Deep Spiking Neural Networks (DSNNs) mimic the human brain's hierarchical structure for effective spatio-temporal information processing.
• Supervised learning for DSNNs at the spike train level is challenging due to complex structures and nonlinear mechanisms.

Purpose of the Study:

• To develop novel supervised learning algorithms for DSNNs.
• To address the weight transport problem inherent in backpropagation (BP) for DSNNs.
• To evaluate the performance and stability of proposed algorithms.

Main Methods:

• Proposed BP-STIP algorithm based on kernel functions and spike trains inner product (STIP) with error backpropagation.
• Introduced FA-STIP and BA-STIP algorithms utilizing feedback alignment (FA) and broadcast alignment (BA) to optimize error feedback.
• Validated algorithms on the MNIST handwritten digit dataset.

Main Results:

• FA-STIP achieved 94.73% classification accuracy, and BP-STIP achieved 95.65% accuracy on MNIST.
• The proposed FA-STIP and BA-STIP algorithms demonstrated improved learning performance and stability over the benchmark BP-STIP.
• Analyzed the impact of different kernel functions on DSNN learning across various network scales.

Conclusions:

• The developed FA-STIP and BA-STIP algorithms offer effective solutions for supervised learning in DSNNs.
• These methods show enhanced performance and stability, outperforming existing benchmark algorithms.
• The study highlights the potential of biologically inspired learning mechanisms for advanced neural network applications.