Modeling Feedback Mechanism In Deep Convolutional Neural Networks

With the development of artificial intelligence,deep neural networks have been widely applied in various fields of human society.In particular,feedforward deep con-volutional neural networks(CNNs)have been a great success in computer vision.How-ever,with the increasing demand for practical application,the visual tasks for CNNs become more and more complex.The simple feedforward architecture of CNNs lim-its them to further meet the requirements.So,it is necessary to develop new mech-anisms of visual information processing,which are more similar to the human visual system.Studies from Cognitive Neuroscience show that the human visual system con-sists of a large number of feedforward connections,feedback connections and lateral connections,and it generally contains much more feedback and lateral connections than feedforward connections.Therefore,it has important research value to model effective feedback mechanism based on feedforward networks.In this thesis,inspired by the feedback mechanism of human visual cortex,we give a deep research on how to build a computational model of feedback mechanism in deep convolutional neural networks.And effective computational models of feedback mechanism are proposed,which fully extend the function of convolution neural networks and have been successfully applied to a number of computer vision tasks.The main research work and contributions are as follows:1.This work theoretically proposes a new feedback mechanism in deep convolu-tional neural networks.The mathematical definition of the feedback mechanism is proposed for deep convolutional neural networks which are only trained for classification task.We formulate the feedback mechanism as an optimization problem.And we propose a basic feedback framework based on convolutional neural networks.2.This work proposes a gradient descent method to solve the feedback optimiza-tion problem.The proposed feedback convolutional neural networks effectively model the mechanism of selective attention in human vision.Experimental results on large-scale datasets verify that the proposed feedback model can effectively capture the interested objects and enhance the object recognition ability of CNNs.3.This work further proposes two new algorithms based on the greedy strategy to solve the feedback optimization problem.We analyze the functional differences of these two algorithms,and propose a novel feedback convolutional neural net-work(feedback CNN)based on the neural pathway pruning and pattern recover-ing algorithms.A large number of experimental results on the tasks of weak su-pervised object localization and semantic segmentation verify the effectiveness of the proposed feedback CNN.Furthermore,our method also provides visual explanations for the working principle of deep convolutional neural networks.4.In order to explore the information processing mechanisms which are more sim-ilar to the human visual system,this work proposes a new computational model of lateral inhibition based on the feedback mechanism and deep convolutional neural networks.By applying this model,both the tasks of selective attention and salient object detection are integrated into CNN classifiers.Extensive exper-imental results verify the effectiveness of the proposed model.In summary,this work formulates the feedback mechanism as an optimization problem in deep convolutional neural networks,and proposes three different solutions.And,while successfully modeling the feedback connections,this work further con-structs lateral connections in the convolutional neural networks,and proposes a visual attention model based on both the feedback mechanism and lateral inhibition.A detailed analysis of the feedback convolutional neural networks based on extensive experiments is given.The results show that the proposed model of feedback mechanism greatly in-creases the flexibility of deep CNNs and extends their function and application scope.The research work in this thesis provides a new insight to implement brain-inspired concepts,and is an important attempt to explore more intelligent visual systems.