Deep Learning Based On Immune-convolutional Neural Network

As a new field of machine learning,amazing achievements have been made both in the theory and application of deep learning recently.Originates from the study of artificial neural network,the main purpose of deep learning is to establish and simulate the neural network of human brain.Besides,the success of deep learning contributes to the fast-growing of the internet either,i.e.providing both a powerful computer resource and a wealth of training data for deep learning.With the powerful ability of self-learning layer by layer,the theoretical study of deep learning with other theory going deeper while the applications expanding faster,however,the scarce and the multi-label property of labeled training data in the non-stationary environment calls for more human involvement which limiting the usage and performance of deep learning to some extent.Therefore,researching on deep learning theory and its application in the non-stationary environment becomes one of the most valuable and challenging hotspots in the field of deep learning.In this paper,research on the theory and application of the deep learning in the background of image processing under the consideration of the characteristics of the big data in non-stationary environment are the main thesis.The main works are arranged as follows:(1)A significant amount of work have been done to dig deep about the theory of deep learning,which focus on the architecture,data argumentation and visualization of deep neural networks.Based on that theory we can provide a solid basis for the following study of the characteristics of the big data in non-stationary environment.At the same time,we mainly study how to enhance the performance of deep learning using transfer learning by initialize the parameters of the deep neural networks in domain with the features in related domain when the labeled training data in domain is scarce while wealthy in related domain,so as to provide theoretical basis for the following modeling of convolutional neural networks based on immune theory.(2)Combining with the multi-label property of the big data in non-stationary environment,a novel multi-label classification model using convolutional neural networks(CNNs)is proposed.Different to the CNNs which concentrated in the background of single-label samples,our model introduce the hidden semantic between different labels of the same sample to the existed CNNs to enhance the performance.Experimental results demonstrate that our model can achieve better classification performance on the multi-label data sets of CIFAR-100 than the CNNs using single label,our model improves the classification performance,by 2.3% increasing for the top-1 accuracy,while 2.7% for the top-5 on average.(3)Aims at realizing the reuse of convolutional neural network architecture by theoretical cross-over study between the theory of deep learning and the immune theory,we proposes a novel convolutional neural network model based on immune theory.The reconstruction of theconvolutional neural network is accomplished by selecting the best antibody depending on the results of affinity between the unknown antigen and the antibody.The unknown antigen is the feature of the unknown image extracted via denoising auto-encoder,and the antibody refers to the feature of the training data which are learned by supervised training using convolutional neural networks,i.e.the convolution kernels.Finally,the model is analyzed in combination with the experimental results,and it provides direction and idea for the follow-up model improvement and research.At last,the paper summarizes the future research content and prospects.