Construction method for network connection adaptive deep convolutional model

Abstract

The invention provides a construction method for a network connection adaptive deep convolutional model. The method comprises the steps that first, weight vectors in a convolutional neural network areorthogonalized; second, connection between layers in the convolutional neural network is deleted according to a norm; and third, an activation function is constructed. According to the method, data relevancy is lowered through weight vector orthogonalization; the connection between layers in the convolutional neural network is deleted based on the p norm, the problem of overfitting in the construction process of the deep convolutional model is effectively lowered, and network connection adaptation is realized; and the activation function is constructed to generalize data information as much as possible on the premise of not changing a parameter scale, and therefore the accuracy of the deep convolutional model is improved.

Description

technical field [0001] The invention relates to the technical field of data processing, in particular to a method for constructing a network connection self-adaptive deep convolution model. Background technique [0002] With the continuous development of information technology and the popularization of electronic equipment, people rely more and more on the Internet, and at the same time leave a large amount of data on the Internet. With the advancement and development of deep learning research, people began to apply deep learning models to solve real-world problems. However, due to the huge amount of data and complex reality, the deep learning model often has many problems during the training process, such as: too many model parameters during the training process, serious data redundancy, and greatly increased training time. Some of these problems come from some imperfections of the traditional neural network. For example, the fully connected layer in the traditional neural...

AI Technical Summary

Problems solved by technology

[0015] The invention provides a method for constructing a network connection self-adaptive deep convolution model, which is used to solve the problem of serious data redundancy in existing deep learning methods, and improves the accuracy and efficiency of deep convolution model construction

Images