SparseMaps: Convolutional networks with sparse feature maps for tiny image classification

Abstract

Abstract Deep convolutional models have been able to get extraordinary results in visual, speech and textual processing domains. Nevertheless, the complexity of manifold of images in data space necessitates the deep networks to have a large number of parameters and subsequently these deep models prone to redundancy and overfitting. One of the most effective methods to tackle this problem is devising special regularization methods in the context of convolutional models. In this paper, after presenting and discussing the most important models in the field, sparse feature maps are proposed and employed as a penalty term in the cost function. This conducts the learning process to construct kernels so that the feature maps at every point are sparse along the depth. A sparse representation is able to adapt to the varying level of information of natural images and can help to extract more independent and informative representations. Also, the DropMaps concept is proposed and employed in the last convolutional layer of the model. This technique applies dropout on the feature maps that causes coincidence of feature maps to be avoided. As is shown in this paper, sparse feature maps and DropMaps can handle the problem of overfitting in large models for tiny images. We have studied the effect of the sparsity rate on the accuracy of the model, and it is observed the accuracy of the test dataset reaches its maximum at a sparsity rate of 0.05. Moreover, by designing appropriate learning rate curves, we were able to obtain ensemble machines with much less cost for training. It is noticeable that the test accuracy is higher than the validation accuracy of the ensemble, indicating that the model has not overfitted. In the input of the model, a random online preprocessing layer is employed for the training phase that helps regularization of the model. Comparing input space and feature space of the model we found that the proposed network is able to successfully separate images of different classes. Finally, testing the proposed model with MNIST dataset has shown that the test set can be classified with accuracy 99.75. The same test with CIFAR 10 dataset attained an accuracy of 94.05.