Efficient Training of Supervised Spiking Neural Network via Accurate Synaptic-Efficiency Adjustment Method.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Spiking neural networks (SNNs) are advanced neural networks excelling in cognitive tasks like pattern recognition.
• SNNs utilize temporal encoding, inspired by the hippocampus, offering superior computational power.
• However, serial processing in temporal encoding significantly hinders SNN training efficiency.

Purpose of the Study:

• To introduce a novel training algorithm for SNNs that enhances efficiency without compromising computational power.
• To address the low learning efficiency associated with traditional temporal encoding in SNNs.
• To develop a method that leverages biological mechanisms for improved SNN training.

Main Methods:

• Proposed the accurate synaptic-efficiency adjustment method, inspired by primate visual selective attention.
• Focused training on target spike times, ignoring irrelevant voltage states to reduce computational load.
• Implemented a cost function based on output neuron voltage difference from the firing threshold.
• Utilized a normalized spike-timing-dependent-plasticity (STDP) learning window for error assignment to synapses.

Main Results:

• The proposed algorithm significantly reduces SNN training time.
• Demonstrated higher learning performance compared to existing SNN training methods.
• Achieved state-of-the-art efficiency in training SNNs with both single and multiple spike inputs.
• Simulations confirmed the algorithm's effectiveness across various input scenarios.

Conclusions:

• The accurate synaptic-efficiency adjustment method offers a breakthrough in SNN training efficiency.
• This biologically inspired approach successfully balances computational power with learning speed.
• The findings pave the way for more practical and efficient applications of SNNs in complex cognitive tasks.