A Dynamic Adaptive Data Truncation Method for Convolutional Neural Network Computing

Abstract

The present invention provides a dynamic self-adaptive data truncation method for convolutional neural network calculation, comprising: expanding m-digit decimal point position data after t-bit image data, expanding m-digit decimal point position data behind t-bit weight data; In the multiplication operation, t-bit image data and t-bit weight data are multiplied to obtain 2*t-bit result data, and two m-bit decimal point position data are added to obtain the decimal point position of 2*t-bit result data, which is recorded as M ; Compress the high-order 0 value in the 2*t-bit result data, the compressed data is intercepted from the high-order to obtain t-bit result data, and the t-bit result data is obtained and intercepted according to the bit width and M of the truncated low-bit data The corresponding decimal point position; splicing the t-digit result data with the corresponding decimal point position data. While ensuring high-order data is retained, as many decimal places as possible are retained, so that the data operation accuracy is as high as possible under a given hardware architecture.

Description

technical field [0001] The invention relates to a data truncation mechanism in a convolutional neural network hardware acceleration process, in particular to a dynamic adaptive data truncation method for convolutional neural network calculation. Background technique [0002] Artificial intelligence is one of the most popular computer sciences at present. As the main way to realize artificial intelligence, deep learning has also achieved profound development. Convolution Neural Network (CNN) is one of the most studied and widely used network structures of artificial neural network structures. It has become one of the research hotspots in many scientific fields, especially in the field of pattern classification. Since CNN avoids the The complex preprocessing of the image can directly input the original image, so it has been more widely used. In recent years, convolutional neural networks have made great achievements in the field of computer vision, and at the same time, convo...

AI Technical Summary

Problems solved by technology

However, for some convolutional layers, the data range is very small, and the required integer bit width is smaller than the reserved bit width, which leads to redundant integer bit width, insufficient decimal bit width, and reduced data accuracy.

[0008] However, the problem with these truncation strategies is that the accuracy of the truncated data is insufficient, and it is even possible that the truncation of high-level data may lead to data truncation errors

Images