Training Deep Spiking Convolutional Neural Networks With STDP-Based Unsupervised Pre-training Followed by Supervised...

Area of Science:

• Neuromorphic Computing
• Artificial Intelligence
• Computational Neuroscience

Background:

• Spiking Neural Networks (SNNs) offer power efficiency and accuracy for cognitive tasks.
• Deep SNNs with multiple layers are needed for complex functional representations.
• Initializing deep SNNs effectively is crucial for performance.

Purpose of the Study:

• To propose a biologically plausible unsupervised pre-training scheme for deep SNNs.
• To improve parameter initialization before supervised optimization.
• To enhance the training efficiency and generalization of SNNs.

Main Methods:

• A two-phase training approach for multi-layer SNNs (convolutional, pooling, fully-connected layers).
• Layer-wise pre-training of convolutional kernels using unsupervised Spike-Timing-Dependent-Plasticity (STDP).
• Fine-tuning of synaptic weights using spike-based supervised gradient descent backpropagation.

Main Results:

• STDP-based pre-training significantly improves robustness and generalization.
• Achieved approximately 2.5x faster training times compared to purely gradient-based methods.
• Demonstrated superior performance on digit recognition tasks.

Conclusions:

• Unsupervised STDP pre-training is an effective strategy for initializing deep SNNs.
• This hybrid approach enhances SNN training efficiency and predictive accuracy.
• The proposed method advances brain-inspired computing for complex pattern recognition.