Improving Deep Learning Models in Weakly Supervised Learning Paradigms

Abstract

Convolutional Neural Networks (CNNs) have provided promising achievements for image classification problems. However, training a CNN model relies on a large number of labeled data. Considering the vast amount of unlabeled data available on the web, it is important to make use of these data in conjunction with a small set of labeled data to train a deep learning model. In this thesis, we aim to develop some methods that can make use of unlabeled data along with partially labeled data to obtain a better generalization performance for CNN models. Moreover, it is also often easier to collect labels that capture only part of the information about the true label of interest. A particularly pertinent example is semantic labels obtained from hashtags attached to images. Such tags are generally easy to gather in large quantities, but tend to only capture certain aspects of the image that the person tagging them focused on. For example, objects are usually organized in a hierarchical structure in which each coarse category (e.g., big cat) corresponds to a super-class of several fine categories (e.g., cheetah, leopard). The objects grouped within the same coarse category, but in different fine categories, usually share a set of global features; however, these objects have distinctive local properties that characterize them at a fine level. Therefore we can tackle the challenge of fine image classification in a weakly supervised fashion, whereby a subset of images is tagged by fine labels (i.e., fine images), while the remaining are tagged by coarse labels(i.e., coarse images).