Abstract
Image classification is an important task in computer vision which aims at classifying images based on their content. Most techniques for this task require a lot of labeled data to train the model which is scarce and expensive. Self-taught learning tackles this problem by using a generic dataset of unlabeled data to train a generative model and use it for feature discovery. Deep learning techniques have been known to perform well for feature discovery. Convolutional deep belief networks with probabilistic max pooling provide a translational invariant hierarchical generative model supporting both top-down and bottom-up inference that perform well on images. In this work, we employ CDBNs for self-taught learning to learn features from images to classify them.
Downloads
Project ProposalSlides
Poster
Report
Code