| Convolutional Neural Network(CNNs)is the focus and hotspot in the field of artificial intelligence at present.There has been numerous work on convolutional neural networks that focus on how to achieve good performance in tasks such as classification and detection.In contrast,the semantic arrangement of CNNs neurons has been largely neglected by current researches.Because it can't significantly improve the accuracy of these tasks.Inspired by both the columnar organizations of neurons in human visual cortex and the semantic network model in psychology,we attempt to produce CNNs with the semantic arrangement property based on image classification in this paper.Here by semantic arrangement,we mean that neighboring neurons are activated by semantically closely related concepts.In practical application,the semantic arrangement property could offer high efficiency for both concept storage and retrieval.In order to realize the semantic arrangement of CNNs neurons,this paper proposes two different approaches.The one is based on self-organizing mapping algorithm and the other is regularization.Besides pointing out,it's the first time in the field of deep learning to do some research about semantic arrangement of neighboring CNNs neurons.The specific work is as follows:First,for image recognition tasks,the neurons in the output layer of the network correspond to the categories of datasets.When no manual annotation of semantic closeness between categories is provided,it is impossible to measure the degree of semantic arrangement of neurons.So we use the weights similarity of CNNs filter kernels as a measure of the semantic arrangement of CNNs neurons.In order to make the neighboring filter kernels have similar weights to each other,a regularization-based approach is firstly proposed,which adds regularization term to cross-entropy loss function to realize the semantic arrangement of CNNs neurons.Second,noticing that the self-organizing mapping algorithm can generate similar weights for neighboring neurons,we propose in our second approach to exploit SOM to make the neighboring filter kernels have similar weights to each other.However,the traditional SOM is for fully-connected 2-layer networks,and we therefore modify it in the context of CNNs.The most important modification is related to convolution.Third,due to the lack of evaluation criteria for semantic arrangement between neurons in convolutional neural networks in current research tasks,we propose three approaches to evaluate its advantages.The first measures the similarity between neighboring filters.The second examines the semantic distances between categories in CNNs and compares them with those in WordNet.The last tests the prediction capability enabled by activation diffusion.Our experimental results demonstrate that,as compared with traditional CNNs,much better semantic arrangement of CNNs neurons has been achieved by our approaches and the semantic distribution between ImageNet categories in resulting CNNs is much more consistent with those in WordNet. |
