Paired competing neurons improving STDP supervised local learning in spiking neural networks

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Machine Learning

Background:

• Spiking Neural Networks (SNNs) offer energy-efficient computation on neuromorphic hardware.
• Unsupervised Spike Timing-Dependent Plasticity (STDP) enables gradient-free learning but is insufficient for classification.
• Supervised learning methods are needed to adapt SNNs for classification tasks.

Purpose of the Study:

• To develop a supervised learning rule for SNN classification layers.
• To enhance SNN classification performance using a novel training architecture.
• To enable efficient and accurate SNN-based image recognition on neuromorphic platforms.

Main Methods:

• Proposed Stabilized Supervised STDP (S2-STDP) with error-modulated weight updates aligning neuron spikes.
• Introduced Paired Competing Neurons (PCN) architecture for neuron specialization and intra-class competition.
• Evaluated S2-STDP and PCN on MNIST, Fashion-MNIST, and CIFAR-10 image recognition datasets.

Main Results:

• S2-STDP outperforms existing supervised STDP learning rules in classification accuracy.
• The PCN architecture significantly enhances S2-STDP performance across various datasets.
• PCN improves S2-STDP effectiveness without introducing additional hyperparameters.

Conclusions:

• S2-STDP provides an effective supervised learning method for SNN classification layers.
• The PCN architecture is a valuable addition for boosting SNN classification capabilities.
• These advancements pave the way for more powerful and energy-efficient SNNs on neuromorphic hardware.