Research On Construction And Application Of Octonion Convolutional Neural Network

In recent years,Real Neural Network(Real NN)has received widespread attention in both academia and industry.The construction,promotion and reasonable interpretation of the network are important research contents of the basic theoretical research in the field of artificial intelligence application.As a classic learning model for deep learning,Real Convolutional Neural Network(Real CNN)has made breakthroughs in the fields of speech recognition,image processing,medical aided diagnosis,etc.,but its network structure usually does not consider the correlation between different convolution kernels.Real Recurrent Neural Network(Real RNN)uses convolution cores to establish the connections and learn their weights to obtain the correlations,but this method greatly increases the difficulty of training and may have poor convergence is prone to over-fitting problems.In order to surmount these shortcomings,we propose a method that can consider the correlation between convolution kernels without learning the correlation by adding connections between the convolution kernels.In this thesis,the idea is to propose a more general framework of neural networks,nameed deep octonion convolutional neural network(DOCNN),which can be regarded as an extension of convolutional neural networks from the complex,quaternion domain to the octonion domain.The contributions of this dissertation are summarized as follows:(1)We describe the main building blocks of DOCNN in which the octonion convolution module,the octonion batch normalization module and the octonion weight initialization module are defined;then explain the effectiveness of deep complex convolution neural network(DCCNN),deep quaternion convolution neural network(DQCNN),and DOCNN from the perspective of multi-task learning.(2)We apply the proposed deep DOCNN to the classification tasks on the database including CIFAR-10 and CIFAR-100.By using the creteria such as the number of floating-point operations(FLOPs)and multiply-accumulates(MACCs)performed per second,running time of the model is evaluated.The results show that,compared with other deep networks,DOCNN can achieve lower error rate with less parameter number and running time.Moreover,when more categories need to be distinguished,its advantages will become more obvious.(3)We use the proposed octonion convolution module for speech classification tasks.Multi-dimensional feature combinations are treated as a single entity which isthe feature input of the octonion convolutional neural network.Experimental results show that the input features combined in this way have stronger expression ability.Compared with other shallow networks,shallow octonion convolutional neural network can still achieve better classification results.