Neural network operational circuit based on digital-analog hybrid neuron

Abstract

The invention discloses a neural network operational circuit based on digital-analog hybrid neurons, and relates to the technical field of neural networks. A multiplication operation circuit is used for carrying out multiplication operation on the quantization weight and the fixed-point quantization activation value, the multiplication operation circuit is realized by adopting N shift registers, wherein each shift register shifts and outputs m bits each time, and the addition operation circuit adopts an analog circuit to carry out wide vector summation operation on the m bits shifted and output by the shift registers in sequence, an AD conversion circuit is used for converting the wide vector summation result output by the convolution circuit into a digital signal, wherein the quantization weight is the exponential power of 2 or 0. The circuit is suitable for a mobile terminal and portable equipment, the precision of the neural network model is improved, the accuracy of the output result of the neural network model is improved, the chip area is reduced, the neural network model can be deployed on the terminal, the operation precision is ensured, the problems that a high-precision neural network is high in power consumption and hardware is difficult to implement are solved while precision improvement is met.

Description

technical field [0001] The invention relates to the technical field of neural networks, in particular to a neural network operation circuit based on a digital-analog mixed neuron. Background technique [0002] At present, with the development of neural network technology, in the fields of image processing and speech recognition, deep neural networks have achieved very good results in the cloud. Based on requirements such as latency, bandwidth, and privacy, it is necessary to push the neural network from the cloud to the terminal, and perform reasoning applications such as keyword detection, face recognition, and image classification on the terminal. However, with the improvement of accuracy, the depth and parameter quantity of the neural network will also increase sharply. The existing technology cannot meet the condition of improving the accuracy while overcoming the problem that the neural network consumes a lot of power and is difficult to implement. This leads to high-pr...

AI Technical Summary

Problems solved by technology

However, with the improvement of accuracy, the depth and parameter quantity of the neural network will also increase sharply. The existing technology cannot meet the condition of improving the accuracy while overcoming the problem that the neural network consumes a lot of power and is difficult to implement. This leads to high-precision Neural networks are difficult to deploy on mobile and portable devices

Images