Spiking Neural Networks Training with Combined Hebbian Rules

The article presents a novel local learning rule for spiking neural networks (SNNs). The developed Supervised Learning method uses a combination of Hebbian rules, so-called spike-time-dependent plasticity and all-time long-term depression. The rules, characteristic of biological neurons, together with lateral inhibition and an adaptive threshold, allow us to find the unique input features for every class of objects. Two series of experiments were carried out with and without an adaptive threshold. The classification accuracy was (74.9 ± 19.8)% in experiments without an adaptive threshold and (80.0 ±13.0)% in experiments with an adaptive threshold. SNNs were implemented in the Brian 2.0 package. Experiments were carried out on a freely available dataset of handwritten digits images MNIST.