Computer Vision News - April 2019
the best human players, deep learning still falls short in efficiently shifting between tasks, and especially when the bot is faced with new surprises that is not presented in the training data. The SRI team is approaching the AI challenges with deep reinforcement learning (RL) problem. In terms of algorithmic techniques, the research team uses memory as a way of addressing the problem. They rely on generative modeling techniques to model memory since storing the experiences doesn’t scale up to the size of the problem or the number of different tasks experienced. They encode memory into separable latent space, and then generate those experiences (e.g. like memory recall) to retrain the deep RL algorithm. They are using the StarCraft2 game simulation environment to provide a lifelong learning training environment. As Jesse explains: “ Our approach to lifelong learning is based on the idea of memory, remembering past experiences, and using those to help in new situations and to re-experience past situations so that you don’t forget about them. Our approach is based on the way that the biological memory systems have been found to replay past experiences in a mammal brain. We’re creating an analogous bio- inspired system that learns how to encode experiences and then reuse them in a process called pseudo- rehearsal which allows us to transfer that knowledge to new tasks and to re- 26 Application Computer Vision News Application
Made with FlippingBook
RkJQdWJsaXNoZXIy NTc3NzU=