Training learning method and system for neural network of paper pattern specification parameter inference model

Abstract

The invention discloses a training learning method and system for a neural network of a paper pattern specification parameter inference model. The method comprises the steps that S1, all connection weights and thresholds are subjected to initialization setting after multiple groups of learning samples are imported into an input layer; S2, one group of learning samples are selected randomly, and input and output of all units in an interlayer are calculated through an interlayer algorithm; S3, output and response of all units of an output layer are calculated through an output layer algorithm; S4, an output layer generalized error of all the units of the output layer is calculated, and an interlayer generalized error of all the units of the interlayer is calculated through an interlayer generalized error algorithm; S5, the connection weight of the output layer is corrected, the output threshold of the output layer is corrected through a correction algorithm, the connection weight of the interlayer is corrected, and the output threshold of the interlayer is corrected through the correction algorithm; and S6, the next group of learning samples are selected randomly, and the step S2 is returned to. According to the training learning method and system, a network is converged step by step through selection of a proper training rate and other methods, and therefore the training learning effect is improved.

Description

technical field [0001] The invention belongs to the field of neural network learning, and in particular relates to a training and learning method and system for a neural network of a pattern specification parameter reasoning model. Background technique [0002] Artificial Neural Network (ANN) is a research hotspot in the field of artificial intelligence since the 1980s [32]. From the perspective of information processing, the human brain neuron network is abstracted, a simple model is established, and different networks are formed according to different connection methods [33]. The BP (Back-Propagation) neural network model is the most widely used in the field of artificial neural network models, and it is also the essence of the field of artificial neural networks. Because the middle layer of the BP neural network is generally one or more layers, the entire network layer is three or more layers. [0003] The neural network needs to be trained and learned before or after i...

AI Technical Summary

Problems solved by technology

[0005] The above solution uses at least one computing host to train the replica deep neural network, so that multiple hosts can be used for asynchronous and parallel deep neural network training, which greatly improves the efficiency of deep neural network training, but there are still areas for improvement. For example, the training method is still not flexible enough, the training rate is still not appropriate, etc.

Images