Automatic Gating Of Attributes In Deep Structure

Deep structure has been widely applied in a large variety of fields for its excellence of representing data.Attributes are a unique type of data descriptions that have been successfully utilized in numerous tasks to enhance performance.However,to introduce attributes into deep structure is complicated and challenging,because different layers in deep structure accommodate features of different abstraction levels,while different attributes may naturally represent the data in different abstraction levels.This demands adaptively and jointly modelling of attributes and deep structure by carefully examining their relationship.Different from existing works that treat attributes straightforwardly as the same as raw data,we can make better use of attributes to improve the performance of deep structure by properly connecting them.Followings are detailed works done in this thesis:1.We propose an energy based model,which can automatically connect attributes to appropriate hidden layers in deep neural network.It can jointly model attributes and deep structure by introducing a learnable gate matrix to control the connections between attributes and hidden layers in deep structure.Based on this model,we also propose two extension models.The first one directly combines the convolution operation in the energy function of the proposed model.The second one stacks convolutional neural network and the proposed model by feeding the output of convolutional neural network as the input of the proposed model.2.We propose a two-stage iterative learning algorithm,which iteratively learns model parameters by clipping gate matrix and learns gate matrix by clipping model parameters and finally reaches convergence.3.We conduct experiments on three real world data sets against six baseline models in various scenarios.All the results show that the proposed model can make better use of attributes to improve performance of deep structure by appropriately connecting attributes to hidden layers.Specifically,the extension model outperforms convolutional neural network,which is commonly accepted as the state-of-art model in image processing,on several data sets.With the application of the proposed model and algorithm,attribute can be better utilized in deep structure to help improve the performance.In this way,by annotating attributes for raw data,many real-word application can obtain better performance.