چكيده به لاتين
Abstract
In this thesis, in the first Caltech-256 dataset is classified using Convolutional networks and at the next step hyper-parameters of the network is optimized to get accuracy. This network consists of 5 convolutional layers 3 fully connected layers. Every Convolutional layers in turn consists of convolution, pooling and nonlinearity operation. The last layer that is one of the fully connected layes is a softmax that is used to classify the images into 256 classes. In order to improve the model, a bag of tricks have been used. The ReLU nonlinearity is used to prevent neurons to be saturated and slowing down updating the parameters. Dropout technique is used to prevent coherence effect of neurons in the same layer. In addition Dropout causes an implicit ensemble classifier to be formed. Pooling with overlapping is used to make the model invariance with respect to location. And finally data augmentation technique is used to prevent over-fitting of millions of parameters of the system. Convolutional layers are connected together so that maximum parallelizability is attained. After constructing the model various hyper parameters, such as receptive fields sizes, number of kernels, number of Convolutional layers, nonlinearity types, and dropout rate are tuned so that maximum accuracy for Caltech-256 data set is obtained. After interpretation of the results, employing slope of 0.15 in negative region of a Leaky-ReLU nonlinearity accuracy of 78 percent gained.
Keywords : Image Classification, Convolutional Networks, Deep Learning, Caltech-256 Dataset, Hyper-parameter Tuning
