Deep Learning

SPLIT: Out-of-distribution Generalization with Granularity Biased Variational Autoenscoder

A variational autoencoder (VAE) model that captures global and local variations in its representation can improve generalization performance in downstream tasks. We propose the Separated Paths for Local and Global Information (SPLIT) framework that modifies standard VAE models, enabling them to disentangle global and local visual features explicitly. We apply our framework to solve several downstream tasks: clustering, unsupervised object recognition, and visual reinforcement learning. The learned representations are found to be useful for these tasks and improve out-of-distribution generalization performances.

Maytus Piriyajitakonkij*, Rujikorn Charakorn*, Sirawaj Itthipuripat, Poramate Manoonpong, and Nat Dilokthanakul
SPLIT: Out-of-distribution Generalization with Granularity Biased Variational Autoenscoder
Dynamics-aware Reinforcement Learning for Playing Tennis Games
Sports are straightforward for humans to play. However, none of the existing robot learning algorithms achieve the human-level abilities to play sports. In this project, we investigate the integration of dynamics prediction and control in AI agents for playing sports to make them more adaptable to dynamic changes. Our contributions are as follows. First, we develop a new learning environment called Tennis2D for developing reinforcement learning (RL) algorithms to play tennis. Tennis2D contains realistic dynamics and has dynamic adaptation challenges for single-agent and multiagent RL. Second, we proposed a new dynamics prediction model called Contrastive Dynamics Model (CDM). The model exploits contrastive representation learning to estimate unknown physical parameters and incorporates estimated parameters to improve prediction accuracy. Lastly, we proposed Dynamics-aware RL (DynARL) that utilises model-based prediction and model-free learning. DynARL is found to improve the generalisation performances of both in-distribution and out-of-distribution dynamics. The DynARL-trained agents are found to be more robust to windy circumstances.
Dynamics-aware Reinforcement Learning for Playing Tennis Games