Method and apparatus for fixed-pointing layer-wise variable precision in convolutional neural network

Abstract

The invention discloses a method and an apparatus for fixed-pointing the layer-wise variable precision in a convolutional neural network. The method comprises the following steps: estimating fixed-pointing configuration input to various layers in the convolutional neural network model respectively in accordance with input network parameters and a value range of input data; based on the acquired fixed-point configuration estimation and the optimal error function, determining the best fixed-point configuration points of the input data and network parameters of various layers and outputting the best fixed-point configuration points; inputting respectively the input data which is subject to fixed-pointing and an input data of an original floating-point number as a first layer in the convolutional neural network and computing the optimal fixed-point configuration point of the output data of the layer, and inputting the output result and an output result of the original first layer floating-point number as a second layer. The rest of the steps can be done in the aforementioned manner until the last layer completes the whole fixed-pointing. The method of the invention guarantees the minimum precision loss of each layer subject to fixed-pointing of the convolutional neural network, can explicitly lower space required by storing network data, and can increase transmitting velocity of network parameters.

Description

technical field [0001] The invention relates to the field of computer technology, in particular to a layer-by-layer variable-precision fixed-point method and device of a convolutional neural network. Background technique [0002] With the continuous development of Internet technology and sensor technology, our era is gradually entering the era of big data. This surge in the amount of data has had a huge impact on people's lives, and many laws that were originally difficult to discover have been discovered with the rise of people's research on big data. At the same time, the skyrocketing amount of data has also challenged the way people process data. Traditional manual analysis and processing of data has become a drop in the bucket. We need to let machines learn to process data by themselves. As a class of algorithms that allow machines to independently identify various features and extract useful information from a vast data set, machine learning algorithms are also produc...

AI Technical Summary

Problems solved by technology

[0003] Among them, the convolutional neural network is an important algorithm in machine learning algorithms. Compared with some traditional algorithms in the field of image processing, it is much more prepared and promotes the development of the computer field. With people's continuous research, the convolutional neural network As the recognition accuracy continues to improve, the computational and storage complexity of its algorithms also increases. The ever-expanding computational complexity and space complexity pose challenges to the performance of computing devices.

[0004] The huge amount of calculation and storage of the convolutional neural network makes it difficult for general embedded computing platforms to perform calculations efficiently, making it difficult for many machine vision applications that require high real-time performance to be implemented on the embedded side.

Images