Attention Driven Dynamic Memory Maps
- Sahni, Himanshu*; Bansal, Shray; Isbell, Charles
- Accepted abstract
[Join poster session]
Poster session from 15:00 to 16:00 EAT and from 20:45 to 21:45 EAT
Obtain the zoom password from ICLR
In order to act in complex, natural environments, biological intelligence has developed attention to collect limited informative observations, a short term memory to store them, and the ability to build live mental models of its surroundings. We mirror this concept in artificial agents by learning to 1) guide an attention mechanism to the most informative parts of the state, 2) efficiently represent state from a sequence of partial observations, and 3) update unobserved parts of the state through learned world models. Key to this approach is a novel short-term memory architecture, the Dynamic Memory Map (DMM), and an adversarially trained attention controller. We demonstrate that our approach is effective in predicting the full state from a sequence of partial observations. We also show that DMMs can be used for control, outperforming baselines in two partially observable reinforcement learning tasks.