Abu Dhabi Machine Learning Season 2 Episode 4
23.03.2022 - Abu Dhabi Machine Learning - ~2 Minutes
When?
- Wednesday, March 23, 2022 from 6:00 PM to 7:00 PM (Abu Dhabi Time)
Where?
- At your home, on zoom. We are looking at hosting in-person events (and broadcasting online on zoom). Any companies or institutions interested in hosting us?
Programme:
Talk 1: Ensuring Safety for Reinforcement Learning
Constrained reinforcement learning (CRL) agents can trade off performance and safety autonomously based on the constrained Markov decision process (CMDP) framework since it decouples safety from reward. Unfortunately, most CRL agents only guarantee safety after the learning phase. In this talk, we present two approaches that can enhance the safety of an RL agent.
The first setting provides a concise abstract model of the safety aspects, a reasonable assumption since a thorough understanding of safety-related matters is a prerequisite for deploying CRL in typical applications. We propose a CRL algorithm that uses this abstract model to learn policies for CMDPs safely, that is, without violating the constraints during the learning phase. In the second setting, the agent learns to behave safely without a goal in a simulated/controlled environment, which allows unsafe interactions and provides the safety signal. Eventually, this agent is deployed in a target task, where it has a goal and safety violations are not allowed anymore. We draw from the transfer learning framework to train a new policy for the target environment without violating the safety constraints using the policy from the initial environment.
Speaker: Thiago is a PostDoc researcher at Radboud University Nijmegen advised by Dr. Nils Jansen. Previously, he was is a Ph.D. candidate within the Algorithmics Group at Delft University of Technology, advised by Dr. Matthijs Spaan. His research interests lie primarily in the automation of sequential decision-making, focusing on reinforcement learning and its safety aspects. He obtained his M.Sc. degree in artificial intelligence from the Instituto de Matemática e Estatística at Universidade de São Paulo under the supervision of Dr. Leliane N. de Barros and a bachelor degree in computer science at the Departamento de Ciência da Computação at Universidade Federal de Lavras.
More information at https://tdsimao.github.io
Talk 2: Diversity in Generative models
Speaker: Tal Kachman
GANs are notoriously difficult to train. In this talk we will present score-based (denoising diffusion) generative models that have the same generative abilities but with a much lower computational cost. We will also discuss the implications of mode collapse in such diffusive based models, the pitfalls and some ideas on how to tackle them.