Supervised Learning With First-to-Spike Decoding in Multilayer Spiking Neural Networks.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Neuromorphic Engineering

Background:

• The brain utilizes spike-based neuronal information processing with diverse temporal coding strategies for efficient sensory representation.
• Applying spike-based computation to real-world challenges, particularly in low-power embedded neuromorphic systems, is a significant goal.

Purpose of the Study:

• To propose a novel supervised learning method for training multilayer spiking neural networks (SNNs).
• To enable SNNs to solve classification problems using a rapid, first-to-spike decoding strategy.
• To explore spike-based encoding strategies for compact data representation in neuromorphic applications.

Main Methods:

• Developed a supervised learning rule for training multilayer SNNs with stochastic hidden neurons and a deterministic output layer.
• Implemented a first-to-spike decoding strategy for classification tasks.
• Investigated various spike-based encoding strategies, including a novel 'scanline encoding' for image data.

Main Results:

• Demonstrated successful classification performance on benchmark datasets, including MNIST.
• The learning rule generalizes well, even with constrained network architectures (few input/hidden neurons).
• Scanline encoding effectively transforms image data into compact spatiotemporal patterns for SNN processing.

Conclusions:

• The proposed learning rule provides a stable and effective method for training SNNs for classification.
• Spike-based encoding strategies, like scanline encoding, are crucial for efficient data representation in neuromorphic systems.
• Optimized network design and dimensionality reduction are vital for practical neuromorphic applications.