A circuit design method for imprecise computational neural networks

Abstract

The invention provides a circuit design method for inaccurate calculation neural network, comprising the following steps: determining the bit width to determine the selection range of the inaccurate multiplier bit width; Point importance ranking is a heuristic random algorithm driven by, for each alternative bit width, a variety of specific circuits are generated, and the various specific circuits are all optional non-exact multiplier structures; independent replacement, the neural network Each accurate multiplier in is independently replaced by an inaccurate multiplier of a certain structure, and all possible solutions obtained form a solution space; traversing the comparison, randomly selecting a part of the solutions in the solution space, and traversing the non-existent solutions under each solution Accurate neural network training and testing results to get the optimal solution. The invention greatly reduces the area and power consumption and improves the calculation speed under the condition that the accuracy of the calculation results is not greatly affected.

Description

technical field [0001] The invention relates to the technical field of artificial neural networks, in particular to a circuit design method for non-accurate calculation neural networks. Background technique [0002] Artificial neural network is an emerging popular algorithm in the field of machine learning, and has achieved good results in many subdivided fields such as face recognition and target detection. However, it relies heavily on hardware performance, and often requires server-level CPUs and GPUs to complete training and recognition, which is costly in terms of size and power consumption. Today, with the miniaturization of smart devices, this characteristic of artificial neural networks makes personal smart devices have to be connected to the Internet all the time, which hinders the further improvement of their speed. However, if the accelerator chip is specially designed for the artificial neural network, its area and power consumption need to be very compressed an...

AI Technical Summary

Problems solved by technology

[0004] However, the imprecise operation unit is not a specific structure, and it is difficult to determine which imprecise structure can be used by the tens of thousands of operation units in the entire neural network to achieve the best effect.

There are still difficulties in the specific application of inaccurate calculation ideas in neural network operations

Images