Supervised learning algorithm based on spike optimization mechanism for multilayer spiking neural networks

Supervised learning is one of the significant research contents in spiking neural networks (SNNs). Aiming at enhancing the performance and efficiency of supervised learning algorithms for multilayer SNNs, this paper proposes a spike optimization mechanism to select optimal presynaptic spikes for computing the change amount of synaptic weights during the learning process. The proposed spike optimization mechanism comprehensively considers the correlation between the desired and actual output spikes of the network. The synaptic weight adjustment is determined by the presynaptic spikes within an optimized time interval, which makes the network output spikes similar to the desired output spikes as much as possible. The spike optimization mechanism is applied to two representative supervised learning algorithms of multilayer SNNs (Multi-STIP and Multi-ReSuMe) to improve their learning performance. The spike train learning results show that the improved algorithms can achieve higher learning accuracy and require fewer learning epochs than the original algorithms. In addition, the spike optimization mechanism can shorten the running time of the algorithms. It indicates that the learning algorithms based on spike optimization are very efficient for learning spatio-temporal spike patterns.

Automated summary

This paper proposes a spike optimization mechanism to enhance the performance and efficiency of supervised learning algorithms in multilayer SNNs. By selecting optimal presynaptic spikes for computing the change amount of synaptic weights, the mechanism considers the correlation between desired and actual output spikes of the network. The application of the mechanism improves learning performance and reduces the running time of the algorithms.