Deep Topology Learning (DeToL)
Deep Learning, i.e. deep neural networks (DNN), have become a key technology in recent years. However, the design of new, problem specific network topologies is still a time and compute intensive process. So far, the design of deep learning solutions for specific applications mostly follows a purely heuristic try and error process based on human expert knowledge and experience. Every network topology needs to be built from a large number of layer types and their configuration. Most layers themselves, as well as the employed training methods, have complex parameter spaces (so-called hyperparameters), whose impact on the final DNN performance is as large as the impact of the network topology itself.
In this project, we aim at facilitating a more efficient topology design process, rendering DNNs accessible to unexperienced users. Within this project, besides the management and organizational tasks, me and my group have the task to design and evaluate suitable graph embeddings that facilitate to explore the network topology space in an efficient way.
DeToL is funded by BMBF. Runtime: October 2018 - September 2021.
- Ying, C., Klein, A., Real, E., Christiansen, E., Murphy, K., & Hutter, F. (2019). NAS-Bench-101: Towards Reproducible Neural Architecture Search. arXiv preprint arXiv:1902.09635.