A biologically plausible supervised learning method for spiking neural networks using the symmetric STDP rule.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Spiking neural networks (SNNs) offer energy efficiency through event-based computation, but supervised training is challenging due to non-differentiable spike activities.
• Existing SNN training methods, such as backpropagation-like and plasticity-based approaches, suffer from limitations including energy inefficiency, non-local dependencies, biological implausibility, or poor performance.
• A gap exists in biologically plausible, high-performance supervised learning methods for SNNs.

Purpose of the Study:

• To propose a novel, biologically plausible spiking neural network (SNN) model for supervised learning (SL).
• To demonstrate the model's effectiveness in achieving high performance on benchmark image recognition tasks.
• To explore the potential for applying the model to neuromorphic hardware and understanding biological learning.

Main Methods:

• Developed a novel SNN model incorporating a biologically plausible symmetric spike-timing dependent plasticity (sym-STDP) rule.
• Integrated sym-STDP with bio-plausible synaptic scaling and dynamic threshold intrinsic plasticity.
• Evaluated the model's performance on the MNIST and Fashion-MNIST datasets, visualizing internal mechanisms using t-distributed stochastic neighbor embedding (t-SNE).

Main Results:

• The proposed SNN model successfully implemented supervised learning, achieving good performance on the MNIST dataset.
• Visualization via t-SNE revealed well-clustered layer activities and synaptic weights, indicating excellent classification ability.
• The model demonstrated robustness and good performance on the more realistic Fashion-MNIST dataset.

Conclusions:

• The novel bio-plausible SNN model effectively performs supervised learning using local, event-driven rules.
• The model's design is suitable for implementation on neuromorphic hardware, enabling online training.
• This work contributes to understanding supervised learning mechanisms at the synaptic level in biological neural systems.