Digital-analog hybrid neuron circuit

Abstract

The invention relates to the field of neural networks, in particular to a digital-analog hybrid neuron circuit, and aims to solve the problem that a large amount of time and energy are consumed duringdata access in the prior art. The circuit comprises an input layer circuit, a plurality of convolution layer circuits and an output layer circuit, wherein the input layer circuit comprises a data memory, a D / A conversion circuit and a weight memory; the data memory is used for storing multi-bit activation values; the D / A conversion circuit is used for converting the activation value of the digital quantity into an analog quantity; the weight memory is used for storing binary weight values in one-to-one correspondence with the activation values; the convolution layer circuits are used for performing multiple convolution operations based on the activation value of the analog quantity and the binary weight value corresponding to the activation value to obtain an analog voltage value; and theoutput layer circuit is used for outputting the digital quantity of the analog voltage value. According to the method, the data value is converted into the analog quantity before the convolution operation, and the activation value participates in the convolution operation with the analog quantity, so the problems of time delay and energy consumption during data access are solved.

Description

technical field [0001] The invention relates to the field of neural networks, in particular to a digital-analog mixed neuron circuit. Background technique [0002] The largest operation in a convolutional neural network is the multiply-add (MAC) operation, and there are millions or even hundreds of millions of MAC operations in a convolutional neural network, which consume a lot of energy. In order to reduce power consumption, a neural network quantization algorithm can be used. However, when the weight is quantized by 1 bit and the activation value is quantized by fixed point (such as 4bit, 8bit, 16bit, etc.), since the activation value is a digital quantity, the addition operation in the convolution needs to use a large amount of capacitance, and the data access operation needs to be expensive. Lots of time and energy. Contents of the invention [0003] The technical problem to be solved by the present invention is to provide a digital-analog mixed neuron circuit for t...

AI Technical Summary

Problems solved by technology

[0003] The technical problem to be solved by the present invention is to provide a digital-analog hybrid neuron circuit for the deficiencies in the prior art. When the weight is quantized by 1 bit and the activation value is quantized by 8 bits, 16 bits or higher precision mbits, the data will be converted before the convolution operation. If the value is converted into an analog quantity, then in multiple convolutional layers of the convolutional neural network, the activation value is always used as an analog quantity to participate in the calculation, which will greatly reduce the number of capacitors in the neuron circuit, reduce the complexity of the neuron circuit, and solve the problem of Calculation delay and energy consumption caused by data transfer between memory and computing unit

Images