Research On The Architecture And Application Of Convolutional Neural Networks

Deep neural networks excel in various tasks such as computer vision and natural language processing.In the field of computer vision,convolutional neural networks have greatly promoted visual tasks such as recognition,detection,and semantic segmentation.Convolutional neural networks(CNN)are used to identify two-dimensional graphics of displacement,scaling,and distortion invariance.Because CNN uses training data for learning,CNN implicitly learns from training data,thus avoiding explicit feature extraction;now deep neural network architecture generally uses information fusion,which explains to a certain extent The layer's network has limited ability to extract valid information,and the subsequent layer information needs to be passed to the front layer.These networks,which incorporate pre-layer information,attribute the improved accuracy to efficient gradient propagation.In this paper,two different network architectures are proposed by designing different information flows.The front layer information flows from the whole to the local(WPNets)or from the local to the whole(PWNets).The convolution of the front layer serves to filter the original information,and to the latter layer,the convolution layer acts as a synergistic enhancement of the original information.Through this kind of competition and cooperation,the network can learn more effective information.Both network architectures have produced good generalization performance on different datasets,proving that both network architectures are very efficient network architectures.Moreover,the number of connections between layers in the neural network is further increased,and information fusion opportunities are further enhanced.Different training methods have a great influence on the generalization ability of neural networks.At the same time,using multiple(two or more)samples to perform convex interpolation training on neural networks can achieve generalization ability better than traditional empirical risk maximization.To explain this phenomenon,several different sample-label mixup interpolation methods were designed to test how the neural network's mixup training method enhances the data of the sample and regularizes the network parameters.Through the proposed network parameter perturbation method,the loss functionof the neural network is visualized,and it is found that the degree of curl of the decision surface of the network trained by the multi-sample convex interpolation method is relatively low.Finally,a spatial mixed sample training method is proposed.The method has a variety of mixup methods,and together with the original channel mixing method,provides more combinations for network training,and the generative adversarial network(GAN)trained using the method is more stable.In order to study whether the neural network can predict multiple samples simultaneously in a forward process,this paper designs a multi-input and multi-output neural network(MSIN).It has been found through experiments that the neural network can not only predict multiple samples at the same time,but also has a relatively high precision.This provides a good solution for the sample domain extension problem.The sample domain extension can be performed on the original network by adding the initial adaptation layer and the final decision layer.The network architecture can also be used to solve the uncertainty problem of neural networks,that is,the neural network should give uncertain predictions for samples that are not in the sample domain.Since the sample domain trained by the neural network is a very small part relative to the whole natural distribution,using the samples outside the sample domain to train together and giving predictions of different sample domains can make the neural network have the predictive ability of uncertainty.We also use this network architecture in the GAN,propose a multi-sample generation confrontation network(MSGAN),and prove in experiments that this network architecture can generate two samples at the same time.