A Novel Unsupervised Spatial–Temporal Learning Mechanism in a Bio-inspired Spiking Neural Network

Bio-inspired computing is a powerful platform that develops intelligent machines based on principles of the behavioral and functional mechanisms of the human nervous system. Such machines can be critical tools in expert systems, speech recognition, pattern recognition, and machine vision. In this study, a retinal model is used as input layer of spiking network to convert image pixels to spike trains. The produced spikes are injected into a spiking neural network (SNN) as a second layer, which structure and functioning is inspired by real neuronal networks (i.e. excitatory and inhibitory neurotransmitters as AMPA and GABA currents and spiking neurons). Similarly, an unsupervised, spatial–temporal, and sparse spike-based learning mechanism based on learning processes in the brain was developed to train the spiking neurons in the output layer for recognizing patterns of MNIST and EMNIST datasets with very high accuracy (above 97%) and CIFAR10 with accuracy 92.9%. The proposed spiking pattern recognition network has higher classification accuracy than previous deep spiking networks and has advantages such as higher convergence speed, unsupervised learning, fewer numbers of hyper-parameters and network layers, and ability to learn with the limited number of training data. Finally, by changing the size and stride of the averaging windows in the visual pathway, we can train the network with only 10% of the training datasets, achieving accuracy similar or higher than state-of-the-art deep learning approaches. The ability to achieve high-performance accuracy in pattern recognition networks despite the limited number of training data is one of the most important challenges of neural networks in artificial intelligence. In summary, the novel bio-inspired neuronal network utilizes spiking trains and learns unsupervised and is capable of recognizing complex patterns, similar in performance to advanced neuronal networks using deep learning, and potentially can be implemented in neuromorphic hardware.