A new approach to image classification by convolutional neural network
Shamim Ibne Shahid, Md. Shahjahan · 2017 3rd International Conference on Electrical Information and Communication Technology (EICT) · 2017
The backpropagation gradient through a convolutional neural network to train the weights of the filters(kernel) is the same as a regular deep network which is to calculate the partial derivative of the error function at the output of fully connected layer with respect to these weights. In this paper we introduce the Fixed Kernel Convolutional Neural Network that skips the training process of the filters between the convolutional layers and exploits the idea of a predefined kernel remaining fixed throughout the training process and a simple feedforward layer that can help acquiring invariances among images having similar kind of patterns at different locations. We compare the performance of the proposed architecture with few other classifying techniques which it outperforms using the Caltech-256 database.