Presented at RSJ2019

At RSJ2019 in Waseda, I presented a following content.

“Parameter Regularization by Integrating Memory Consolidation and Sparsification for Continual Learning”

In this presentation, I proposed a new regularization method for continual learning to mitigate catastrophic forgetting. The conventional method, so-called EWC, has insufficient room for learning new tasks since all the parameters are consolidated more or less. The proposal, named EWCS, has an additional L1 regularization term for sparseness (i.e., initialization), and consequently, it outperformed the number of tasks incrementally learned.

In addition, the following two papers were presented by corresponding students.

“Impedance Control based on Human Motion Estimation for Locomotion Support through Physical Interaction”

“Multi-Agent Reinforcement Learning based on Estimation of State-dependent Interests”