Transfer Learning

MULTIPOLAR, a new transfer RL method presented in IJCAI’20

How to get your models working easily in new environments?

Many modern AI products are powered by cutting-edge machine learning (ML) technologies, such as face detection and language translation apps installed on our smartphones. The development of such AI products typically requires large-scale and diverse data to train the ML models driving them. Nevertheless, collecting data on a large scale is not always easy and moreover, trained models may not work well in a new environment where data is observed under different conditions.

At OMRON SINIC X, we are doing fundamental research on machine learning techniques to make ML models work easily in new environments. Our research interests include transfer learning, unsupervised domain adaptation, and their applications to robotic assembly tasks.

Selected publications

Noisy Universal Domain Adaptation (CVPR2021)

MULTIPOLAR: Multi-Source Policy Aggregation for Transfer Reinforcement Learning between Diverse…

Adaptive Distillation for Decentralized Learning from Heterogeneous Clients (ICPR’20)

TRANS-AM: Transfer Learning by Aggregating Dynamics Models for Soft Robotic Assembly