[11] Jacob Andreas - Learning from Language
The Thesis Review
00:00
How to Automate Machine Translation in a Multi-Agent System
Sine Sookbatar had a paper at Noreps I guess a year or two before I did all of this work showing that you could learn these basically set up a multi-agent system. It was just structured like an RNN with a weird sparse connectivity pattern and learn communication policies. Each agent would send some vector valued message to other agents that somehow was able to communicate what that agent was about to do in the future. So then how do you do machine translation if you don't have access to parallel data but you do have access to something about the world and something about the meaning that these agents might be trying to communicate? That's part of my undergrad thesis is okay how
Transcript
Play full episode
