Training method of convolutional pulse neural network based on reweighted membrane voltage

Abstract

The invention discloses a training method for a convolutional pulse neural network based on a reweighted membrane voltage, and the method comprises the steps: obtaining an input image, carrying out the preprocessing, and obtaining a pulse sequence; constructing spiking neurons of the convolutional spiking neural network based on the reweighted membrane voltage; setting the number of network layers of the convolutional pulse neural network according to the scale of the input image; carrying out normalization processing on the input stimulation of the spiking neurons; constructing a loss function and training the convolutional pulse neural network based on a space-time back propagation algorithm; and inputting the pulse sequence into the trained convolutional pulse neural network to obtain an output result. According to the method, the gradient of the output of the spiking neurons relative to the membrane voltage can be directly calculated, the gradient can be dynamically adjusted according to the accumulated membrane voltage value, and the problem of gradient mismatching in the SNN training process is solved. The training method of the convolutional spiking neural network based on the reweighted membrane voltage can be widely applied to the field of spiking neural networks.

Description

technical field [0001] The invention belongs to the field of spiking neural networks, in particular to a training method for convolution spiking neural networks based on reweighted membrane voltage. Background technique [0002] Spiking Neural Network (SNN) is a neural network model based on imitating the firing mechanism of biological neurons. Due to its ability to efficiently process discrete spatio-temporal events, it has broad application scenarios on low-power devices. At present, the methods of learning SNN can be summarized into two categories. The first method is to convert the pre-trained artificial neural network into the corresponding SNN, and the second method is to train the SNN directly based on the training data. Compared with other methods, the SNN obtained by the first method can obtain better classification accuracy, which is mainly due to the good generalization performance of the artificial neural network (Artificial Neural Network, ANN) obtained by pre-...

AI Technical Summary

Problems solved by technology

However, the good performance of this type of method on classification tasks is based on the huge inference time step. In other words, the first type of method needs to process the input signal in a large time window to obtain good classification accuracy. Inference efficiency is low, and it has certain limitations on low-power hardware

For direct training of SNN, Spike Timing Dependent Plasticity (Spike Timing Dependent Plasticity, STDP), the synaptic weight is adjusted according to the timing relationship of the pre-synaptic pulse arrival, if the pre-synaptic neuron generates a pulse earlier than the post-synaptic neuron, it means There is a causal relationship between the pre- and post-synaptic neurons, and the corresponding synaptic weight increases, and vice versa, the weight value decreases, but the performance of the SNN on the classification task is not satisfactory

The spike-based backpropagation algorithm is another effective method for directly training SNNs. However, this type of method usually needs to estimate the proximal gradient of the output of the spike neuron with respect to the input, which increases the complexity of the algorithm design to a certain extent.

Images