High precision neural network engineering method based on look-up table computation

Abstract

The invention discloses a high-precision neural network engineering method based on look-up table calculation, belonging to the technical field of artificial intelligence neural network depth learning. The method comprises the following steps: taking the output precision as an evaluation index, adopting an adaptive search strategy to sparse the neural network; the weights of the neural network being quantified by nonlinear method, and the data are indexed and quantified. Table-lookup method is used to realize the forward computation of neural network quickly, the invention adopts the sparse rate automatic distribution strategy to achieve the maximum possible reserved network precision under the condition of the set sparse rate, Using the strategy of nonlinear quantization of metric parameters and data index quantization, At the same time, the quantization width is reduced to 8 bits or less, and the computational accuracy is kept very high. Finally, a high-precision and fast computational table of 256 x 256 bits is constructed, which can accelerate the forward process and reduce the computational accuracy without losing, and the computational accuracy is high and the computational speed is fast.

Description

technical field [0001] The invention relates to a neural network engineering method, and more specifically, relates to a high-precision neural network engineering method based on table lookup calculation, which belongs to the technical field of artificial intelligence neural network deep learning. Background technique [0002] Artificial Neural Network (ANN) is a computing model. It is a research hotspot in the field of artificial intelligence since the 1980s. It is mainly composed of a large number of nodes (called neurons connected to each other, each node represents A specific output function, called the activation function, each connection between two nodes represents a weighted value for the signal passing through the connection, called weight, which is equivalent to the memory of the artificial neural network. Network The output of the network depends on the connection method of the network, the weight value and the activation function are different. [0003] A typica...

AI Technical Summary

Problems solved by technology

[0009] The purpose of the present invention is to solve the problem of network precision loss in the traditional engineering process, and to provide a query-based system with technical characteristics such

Images