Method for constructing deep convolution neural network model

Abstract

The invention provides a method for constructing a deep convolution neural network model, belonging to the field of mode recognition and machine learning. The method comprises the steps of (1) initializing a convolution neural network model, (2) carrying out end-to-end global extension learning on a network until the average error of a convolution neural network system reaches a preset expected value, (3) after global extension, using a cross validation sample to evaluate network performance, and carrying out local extension learning on the network if a recognition rate does not reach the expected value, (4) adding a new incremental end-to-end extension learning branch, realizing the incremental expansion learning of a network structure, and finally realizing the model construction of the deep convolution neural network. According to the method, a neural element can be added according to needs according to a condition of participating in training samples, the extension expansion and incremental expansion of the network structure are realized, the relevance between a data sample and the network model is enhanced, and the network structure adaptive incremental learning is realized.

Description

technical field [0001] A method for constructing a deep convolutional neural network model belongs to the field of pattern recognition and machine learning. Background technique [0002] Convolutional neural network is a special deep neural network model. Its particularity is reflected in two aspects. On the one hand, the connection between its neurons is not fully connected. On the other hand, the connection between some neurons in the same layer The weights of the connections between are shared (ie the same). Its non-fully connected and weight-sharing network structure makes it more similar to biological neural networks, reducing the complexity of the network model and reducing the number of weights. The convolutional neural network is a multi-layer perceptron specially designed for recognizing two-dimensional images. It has good fault tolerance, parallel processing ability and self-learning ability, and can handle complex situations such as complex environmental informat...

AI Technical Summary

Problems solved by technology

[0005] But at present, the traditional convolutional neural network structure in the existing technology has shortcomings such as over-reliance on human experience, difficulty in determining the depth of the network and the number of feature maps, and difficult training, and the correlation between data samples and network models is weak.

Images