Representation learning with Less Label and Imperfect Data
Deep learning has attracted tremendous attention from researchers in various fields of information engineering such as AI, computer vision, and language processing. The power of deep learning stems from the ability to learn representations optimized for a specific task, as opposed to relying on hand-crafted features. To yield favorable results, deep models often require a large number of annotated examples for training. However, the data annotating process is expensive, prone to noisy information and human errors, and time-consuming. Moreover, in many applications (such as in medical fields), this process requires domain knowledge and expertise, therefore, often unable to produce a sufficient number of labels for deep networks to flourish. In this work, we develop practical tools to improve the prediction performance of deep neural networks, utilizing less label and imperfect data. In the first part of the thesis we develop the theory to estimate deep neural network prediction uncertainty which measures what the model does not know due to the lack of training data. We tie approximate inference in Bayesian models to DropConnect and other stochastic regularisation techniques and assess the approximations empirically. We further demonstrate the tools’ practicality by making use of the suggested techniques in image processing, natural scene understanding, and medical diagnostics. We exploit Capsule Networks (CapsNets), an alternative proposed to address some of the fundamental issues with training convolutional neural networks (CNNs). We propose novel connectivity techniques and routing mechanisms to extend the use of CapsNets to large-scale, high-dimensional datasets. Our experimental results on several image classification datasets demonstrate that CapsNets compare favorably to CNNs when the training size is large, but significantly outperform CNNs on small size datasets. In the final part of the thesis, we propose a memory-augmented capsule network (MEMCAPS) for the rapid adaptation of computer-aided diagnosis models to new domains. It consists of a CapsNet that is meant to extract compact features from some high-dimensional input, and a memory-augmented task network meant to exploit its stored knowledge from the target domains. Our observations show that MEMCAPS is able to efficiently adapt to unseen domains using only a few annotated samples.