We can now measure a priori how well a deep learning model will parallelize during training. Super pumped that this work uses Dota as a research platform:
Announcing the first finding from our "Science of AI" effort: a simple statistical metric that lets us predict the parallelizability of neural network training on a wide range of tasks: blog.openai.com/science-of-a…
Dec 14, 2018 · 5:34 PM UTC
14
2
106