Title: Applying Genetic Evolution Techniques to the Training of Deep Neural Networks
Description: Deep learning has improved the state of the art across a range of digital content processing tasks. However, the standard algorithm for training neural, the backpropagation algorithm can encounter different types of challenges depending on the network architecture that it used to train. This project will focus on developing novel training algorithms for deep neural networks that can be shown to improve on backpropagation in terms of either final model accuracy, or in terms of computational considerations (such as training time and/or data usage). Furthermore, these training algorithms will be tested across are range of different use cases (e.g., image processing, natural language processing) and network architectures so as to validate the general usefulness of the approach. The initial approach taken to develop these novel training algorithms will be to explore the use of genetic search algorithms.