Width learning system quantification method

Abstract

The invention discloses a width learning system quantification method, and belongs to the technical field of neural network quantification methods. The method comprises the following steps: firstly, obtaining input and output data, secondly, providing a selection standard of quantized weights, then, obtaining an optimal model after random weights are quantized in combination with a search method, quantizing output weights, and finally, obtaining a lightweight width learning system model. According to the method, a dynamic weight quantization mode is used, and the maximum bit which can be selected during weight quantization is set in the training process, so that an integer weight which is optimally matched with the floating point weight is dynamically selected; and finally, grouping quantization and floating point weight retraining of output weights are performed alternately until all weights are quantized and the model is converged at the same time. According to the invention, a lightweight width learning system can be obtained while the performance of the model is maintained or even improved, so that the lightweight width learning system can be more easily deployed on edge applications such as a mobile terminal.

Description

technical field [0001] The invention relates to a neural network quantization method, in particular to a width learning system quantization method, which belongs to the technical field of model compression. Background technique [0002] At present, deep neural networks are very powerful in discovering complex structures in high-dimensional data, but all parameters need to be adjusted by backpropagation, which requires a lot of time; in addition, when the input data changes, the entire network needs to be retrained , which again consumes a lot of time. The proposed efficient discriminative learning algorithm of the width learning system and its variants can solve this problem well. [0003] The breadth learning system has two cores: one is the feature map layer that can extract features and the enhancement layer that increases the nonlinear capability of the network, both of which are connected to the output layer at the same time; the other is that BLS can pass the augmenta...

AI Technical Summary

Problems solved by technology

However, due to the large amount of data interaction in intelligent tasks, the width learning system requires more and more nodes and parameters, which will lead to a larger amount of calculation required by the model, occupying more storage space and consuming more energy. These problems Limits the deployment of width learning systems on edge applications such as embedded

Images