Testing the Quality of Synthetic Data With a Self-Training Approach by Google DeepMind

Hello, fellow human! Sophia here. Wanted to share about a couple of virtual talks we are hosting in the upcoming weeks.

• Imagine virtual worlds where agents have to fight for resources to stay alive, mine materials to get an edge, level up their fighting styles and equipment, try different professions, and trade based on market demand.This is not an MMO game – it's a serious reinforcement learning (RL) research conducted by MIT researchers and their collaborators called Neural MMO 2.0, a massively multi-agent environment for reinforcement learning research.This talk presented by David Bloomin will describe the environment, and the engineering decisions that make it performant.Learn the details of the talk and register here.

• On March 14th we will be hosting a talk by Avi Singh, a Research Scientist at Google DeepMind. He will share about his new research work on ways to test the quality of model-generated data using a self-training approach.Avi and his collaborators investigate a simple self-training method based on expectation-maximization which they called ReSTEM. They generated samples from the model and filtered them using binary feedback. Then fine-tuned the model on these samples and repeated this process a few times.Avi will share the details of this work and the results.Learn the details of the talk and register here.

• I'm also sharing a video recording of the previous virtual talk with Xavier Puig, a Research Scientist from Meta. He introduced Habitat3.0, a simulation platform for studying collaborative human-robot tasks in home environments and how those tasks were transferred to a real-world environment.Watch Xavier’s talk on our YouTube channel.