Fig. 3

Neural Network for unsupervised learning. a Each input digit contains 28 × 28 = 784 pixels and has been cropped and reduced to 397 pixels. The neural network has 397 input neurons with a bias term and 500 output neurons. Input spike trains of input neurons are generated according to pixel density (from 0 to 1) and then fed to the neural network. Synaptic devices represent weights in the network. Top (before training): Random spike activity from representative 10 out of 500 output neurons before learning. Bottom (after training): Output spike trains after learning show coordinated selective firing activity as a result of unsupervised learning of digits. b Spike-timing-dependent plasticity (STDP) rule showing the 10 ms window for an post-pre spike time difference (t_post − t_pre) that determines whether a long-term potentiation (LTP) or a long-term depression (LTD) update is performed. If the firing time of an output neuron (t_post) is within 10 ms of the firing time of an input neuron (t_pre), the weight (synapse) between this input–output neuron pair is updated via LTP. Otherwise, the weight is updated via LTD. c The LTP update is an exponentially decaying function that depends on the current weight, and the LTD update is a constant. The exponential LTP update depending on the current weight keeps the weight values within the range [−1, 1]. d Recognition accuracy vs. number of training epochs. Three epochs are used in our network training. e Recognition accuracy vs. neuron number. Recognition accuracy does not have noticeable increase when number of output neurons is larger than 500. Therefore, 500 output neurons are used in our network model