Growing Artificial Neural Networks Novel approaches to Deep Learning for Image Analysis and Particle Tracking

Abstract

Deep-learning has recently emerged as one of the most successful methods for analyzing large amounts of data and constructing models from it. It has revolutionized the field of image analysis and the algorithms are now being employed in research field outside of computer science. The methods do however suffer from several drawbacks such as large computational costs. In this thesis alternative methods for training the underlying networks are evaluated. These methods are based on gradually growing networks during training using layer-by-layer training as well as increasing network width. These training methods lend themselves to easily implementing networks of tune-able size allowing for tradeoff between high accuracy and fast execution or the construction of modular networks in which one can chose to execute only a small part of the network to get a very fast prediction at the cost of some accuracy. The layer-by-layer method is applied to multiple different image analysis tasks and the performance is evaluated and compared to that of regular training. Both the layer by layer training and the breadth training are comparable to normal training in performance. The modular nature of the networks make them suitable for applications within multi-particle tracking.