List of Projects – Autumn 2019
Theory of reinforcement learning (two independent projects)
For more information, interested students should contact:
– Johanni Brea
– Alireza Modirshanechi and Martin Barry
Effects of weight space symmetries on deep neural network training
Despite the recent success of deep learning, many of its theoretical aspects are yet to be understood. One open question is how deep networks can achieve high accuracies in reasonable training time, given the complex, highly non-convex loss function. In this project we want to explore how weight-space symmetry in multi-layer neural networks could influence training speed both for small and over-parametrized networks. For this, we will start with investigating the weight development during network training. Dependent on our findings there are multiple directions to extend the project, e.g. proposing novel network architectures.
The candidate should have experience in coding (python or julia). Experience in neural-network training and background in theory (i.e. deep learning & matrix algebra) are desirable.
Interested students should send grades and CV to Bernd Illing