Research Scientist @GoogleDeepMind | Agents and Code in Gemini
London, England
Joined March 2012
- Tweets528
- Following7K
- Followers1.7K
- Likes3.1K
Interested in learning more about training agents that can communicate with humans using natural language, stop by at our poster between 4.30pm-6pm GMT at Poster Session-1 (Spot B2) #NeurIPS2021
neurips.cc/virtual/2021/post…
@DeepMind @NeurIPSConf
Tad late to the party but excited to share our recent work on training generalizable agents that can coordinate with unseen/human partners using a dynamic population-based meta-learning approach, accepted at #NeurIPS2021
Grateful to @aggielaz and @sharky6000 for the guidance! 1/7
We show this on two cooperative games- a Lewis signaling game (encoded as symbolic objects) and a more realistic referential game with real images and English captions.
We further show improvements using a population-based approach to S2P. (4/4)
...with the goal of using as few human samples as possible and achieving high reward on the task.
Our findings indicate that it's always best to use all the human data for pretraining before doing self-play, rather than performing self-play first and then fine-tuning later. (3/4)
We introduce a class of learning algorithms called *Supervised self-play* (S2P) that explores systematic ways to combine supervised learning (using a limited amount of human language data) and self-play (agents maximizing reward in a multi-agent setup)... (2/4)
Girls are unstoppable. Watch these girls conquer fear in @always’ new #LikeAGirl PSA: youtu.be/P_MhsbRiFyI