Learning Transferable Representations for Non-stationary Environments

For intelligent agents to become fully autonomous, they need to perceive and adapt to the changes in environmental dynamics. In addition, they need to devise a strategy to acquire new knowledge while retaining the past learned ones. Humans can acquire, retain and transfer knowledge over their lifespan. In a similar vein, intelligent agents are becoming capable of acquiring and transferring knowledge but not retaining it. Towards reaching these goals, we have proposed algorithms that address multiple aspects of machine intelligence, from robot perception, allowing robots to accurately model human intent and predict human motion, to knowledge retention, allowing robots to retain past knowledge without forgetting. Our proposed algorithms have attained state-of-the-art performances for robot perception and overcoming catastrophic forgetting in perception-based tasks. Our current and ongoing work builds upon our completed works to explore knowledge retention in more challenging domains, particularly robot control, and investigate multi-agent collaboration as a precursor for human-robot collaboration.