Neuroscience-Inspired Computational Recognition And Attention Models

Inspirations from neuroscience are useful for computer vision.Human visual systems have excellent visual processing ability.They can fast compress and select from the enormous visual input,effectively represent the information with hierarchical pathway,and use various complex neural mechanisms to adapt the environment.During its history of tens of years' developments,computer vision has gained many inspirations from neuroscience.Some computer models have close relationships with neuroscience findings on visual systems,such as visual feature designing and neural receptive fields,convolutional neural networks and primary visual cortex and hierarchical pathway,and saliency models and visual search experiments.The motivation of this thesis is to propose new computer vision models based on neural inspirations.The content of the thesis can be divided into two parts,corresponding to the two fundamental visual functions including recognition and attention,respectively.In the first part,we propose a hierarchical recurrent neural network model for recognition,and investigate its performance in image classification,scene labeling and EEG signal recognition.The inspiration comes from the fact that recurrent synapses extensively exist in the neural system.In the second part,we investigate the effect of hierarchical feature representations on attention and saliency,and proposes attention and saliency models based on higher-level features.The inspiration comes from the results of some neuroscience experiments on visual attention.In this thesis we extract useful structures and principles from neuroscience findings,propose new computer models based on them,and achieve excellent performance in several tasks.There are two innovations as follow:We propose a novel deep recurrent network model,recurrent convolutional neural network,and apply the model in several tasks.For image classification,the model outperforms other state-of-the-art models and use fewer parameters.For scene labeling,multi-scale recurrent convolutional neural network processes the task in an end-to-end way,and achieves the state-of-the-art performance in both accuracy and speed.We also extend the model from image processing to one-dimensional sequence processing,and successfully apply it to EEG signal recognition.The model achieves excellent performance in an EEG data science competition.Inspired by related neuroscience findings,we investigate attention models based on hierarchical feature representations.Using feature selection as the tool,we analyze the contributions of different features to eye fixation prediction,and construct a state-of-the-art saliency model based on a small number of selected features.We propose two attention models based on higher-level features,and the models significantly outperform traditional low-level models for eye fixation prediction.