Computing Equipment and Electronics

Abstract

The invention discloses a computing device and an electronic device. The computing device includes several computing modules, each of which includes several neuron clusters, and the computing modules are configured as follows: weighted by a first weight matrix, the the input pulse train of the block is projected to the first neuron population through polysynaptic projections, wherein the polysynaptic projections have at least two different synaptic time constants and the two different synaptic time constants are both positive, Or at least two different synaptic transmission delays. In order to realize time-domain convolution in the spiking neural network with low hardware resource consumption, the present invention discloses the multi-synaptic projection technology with different synaptic time constants, and further proposes a method based on residual connection, jumping Waveform-aware spiking neural networks for time-domain signal processing characterized by connections etc. Through these technical means, the performance gap between the SNN and the ANN is bridged, and the SNN whose performance reaches or approaches that of the ANN is obtained.

Description

technical field [0001] The invention relates to a computing device and an electronic device, in particular to a training device, a chip and an electronic device configured with a pulse neural network SNN or an artificial neural network ANN. Background technique [0002] The traditional artificial neural network (Artificial Neural Network, ANN) belongs to the second-generation neural network, and the deep convolutional neural network (existing technology 1) as the representative has led the development of artificial intelligence in the past ten years. [0003] Prior art 1: Krizhevsky A, Sutskever I, Hinton G E. Imagenet classification with deep convolutional neural networks[J]. Advances in neuralinformation processing systems, 2012, 25: 1097-1105. [0004] The second-generation neural network generally takes the pursuit of accuracy as the primary goal (high performance model), so it usually has the characteristics of high energy consumption and high storage space consumption,...

AI Technical Summary

Problems solved by technology

[0015] Temporal convolution is a common basic information processing method in neural networks, such as the aforementioned WaveNet, which is easy to implement in ANN that relies on von Neumann architecture hardware, but it can be implemented at low cost in SNN not easy

Images