Layer increase and decrease deep learning neural network training method, system, medium and equipment

Abstract

The invention discloses a training method, system, medium and equipment for a deep learning neural network with layer number increase or decrease. The method includes: inputting training input data into the current deep learning neural network, and obtaining the first output through the calculation of the current deep learning neural network data; judge whether the first output data is the same as the expected output data; if it does not meet the first preset condition, add a hidden layer before the classifier in the current deep learning neural network; otherwise, input the test input data into the current deep learning neural network Network, the second output data is obtained through the calculation of the deep learning neural network; it is judged whether the second output data is the same as the real result data; if the second preset condition is not met, the previous hidden layer of the classifier in the current deep learning neural network is Delete; otherwise, output the current deep learning neural network. When the present invention can achieve sufficient fitting, the top-level concept is just enough to fully fit the output data.

Description

technical field [0001] The present invention relates to a deep learning neural network training method, in particular to a layer increase and decrease deep learning neural network training method, system, medium and equipment, belonging to the field of neural network training. Background technique [0002] Existing deep learning technology can obtain output labels through input data (such as obtaining the ID card number of the person through the avatar, or obtaining the ID card number of the person through voice). Supervised training (such as avatars with ID number labels, and speech with ID number labels). [0003] However, the top-down supervised training of existing deep learning techniques either only adjusts the network weights between the output layer and the hidden layer, or adjusts the network weights of all layers. When there are more categories of top-level concepts than labels, if only the weights of the classifier network between the output layer and the hidden ...

AI Technical Summary

Problems solved by technology

[0003] However, the top-down supervised training of existing deep learning techniques either only adjusts the network weights between the output layer and the hidden layer, or adjusts the network weights of all layers

When there are more categories of top-level concepts than labels, if only the weights of the classifier network between the output layer and the hidden layer are adjusted, and if the network structure of the classifier is relatively simple, the result of repeatedly adjusting the network parameters of the classifier is often in line with This output label cannot match that output label, which means that a sufficient fit cannot be achieved

If the network structure of the classifier is designed to be very complex, such as using a hierarchically complex BP neural network as the classifier, then there will be over-fitting, which will cause some key features to be discarded during the fitting, so that the sample The classification result is completely correct, but it will be found to be wrong when it is applied

[0004] It can be seen that only supervised training on the level between the output layer and the hidden layer will either fail to fully fit or overfit, which will cause deep learning to fail when applied.

If the network weights of all layers are adjusted, the cognitive weights and generation weights in the hidden layer will be destroyed, so that the concepts and scenes obtained after adjustment are no longer completely derived from the characteristics and scenes of the input data, but for the output labels The features and scenes that are distorted by the needs of the system will also appear overfitting, so that the classification result is completely correct for the sample, but it will be found to be wrong when it is applied.

Images