Skip to content. | Skip to navigation

Sections
Personal tools
You are here: Home Theses Model Training through Curiosity-based Latent Space Exploration on Decentralized Data

Contact

Prof. Dr. S. Decker
RWTH Aachen
Informatik 5
Ahornstr. 55
D-52056 Aachen
Tel +49/241/8021501
Fax +49/241/8022321

How to find us

Annual Reports

Disclaimer

Webmaster

 

 

Model Training through Curiosity-based Latent Space Exploration on Decentralized Data

Thesis type
  • Master
Status Finished
Submitted in 2021
Supervisor(s)
Advisor(s)

Recent developments in privacy rights have highlighted the need for methods capable of training deep learning models on non IID data. In this thesis, we present an approach to train deep learning models on distributed datasets without gathering the data in a central location. We introduce an experimental setup to simulate such distributed datasets with a non-distributed one and show our ndings on MNIST, Cifar-10 and the ISIC challenge. Current methods based on generative models su er from mode collapse when trained on non IID data. As such we develop an approach based on concepts from Distributed Analytics, Active Learning and Reinforcement Learning. Using the idea of measuring informativeness to select samples from Active Learning and the Curiosity Module from Reinforcement Learning to measure said informativeness, we show that curiosity can be used to improve sample eciency and prediction performance. We also show that models trained without our method fail to produce any usable results on data which is both non IID and has a high class imbalance.

Document Actions