Deep Learning in Highly-Correlated Signals with Low Data Availability

Deep learning techniques have revolutionized the field of machine learning, allowing for a paradigm shift in the state of the art in many domains. In particular, deep learning techniques have found success in domains with high spatial or temporal correlation, such as images, video, and audio. However, these advances have relied on the availability of large amounts of data. The success of deep learning in similar domains with high spatio-temporal correlation but lower data availability, such as imaging in manufacturing and neuroimaging, has not been as impressive. Given the inherent spatio-temporal biases in convolutional and recurrent techniques, it is reasonable to believe that deep learning neural network techniques should be able to provide benefit even without the large amount of data seen in the more traditional domains. This dissertation explores a variety of frameworks and techniques to allow for the successful application of deep learning in these low-availability, high-correlation domains. We explore the use of a fully convolutional network for image analysis in manufacturing engineering, a method of pairing signals and determining whether they share a class or not in electroencephalography, and an examination of the use of averaging both as data augmentation in training and for the final signals classified at test time in electroencephalography