Facebook Speeds Up AI Training by Culling the Weak

Facebook engineers have accelerated the training of artificial intelligence (AI) agents by eliminating laggards. The company created the Habitat series of photorealistic three-dimensional simulations to teach AI point-to-point navigation strategies—an essential capability for a practical "embodied AI" or robot.

Previous training systems wasted time waiting for slower agents to catch up, and the researchers created the Decentralized Distributed Proximal Policy Optimization (DD-PPO) system to eliminate these inefficient agents before they complete the task.

Whatever data is accumulated by these laggards is fed into the collective dataset. The DD-PPO system trained agents that navigated virtual environments from starting point to end goal with 99.9% reliability and fewer errors.

From TechCrunch 
View Full Article

Abstracts Copyright © 2020 SmithBucklin, Washington, DC, USA