( events)   Timezone:  
Sat Dec 09 08:00 AM -- 06:30 PM (PST) @ 101 B
Deep Learning at Supercomputer Scale
Erich Elsen · Danijar Hafner · Zak Stone · Brennan Saeta

Workshop Home Page

Five years ago, it took more than a month to train a state-of-the-art image recognition model on the ImageNet dataset. Earlier this year, Facebook demonstrated that such a model could be trained in an hour. However, if we could parallelize this training problem across the world’s fastest supercomputers (~100 PFlops), it would be possible to train the same model in under a minute. This workshop is about closing that gap: how can we turn months into minutes and increase the productivity of machine learning researchers everywhere?

This one-day workshop will facilitate active debate and interaction across many different disciplines. The conversation will range from algorithms to infrastructure to silicon, with invited speakers from Cerebras, DeepMind, Facebook, Google, OpenAI, and other organizations. When should synchronous training be preferred over asynchronous training? Are large batch sizes the key to reach supercomputer scale, or is it possible to fully utilize a supercomputer at batch size one? How important is sparsity in enabling us to scale? Should sparsity patterns be structured or unstructured? To what extent do we expect to customize model architectures for particular problem domains, and to what extent can a “single model architecture” deliver state-of-the-art results across many different domains? How can new hardware architectures unlock even higher real-world training performance?

Our goal is bring people who are trying to answer any of these questions together in hopes that cross pollination will accelerate progress towards deep learning at true supercomputer scale.

Generalization Gap (Presentation)
Closing the Generalization Gap (Presentation)
Don't Decay the Learning Rate, Increase the Batch Size (Presentation)
ImageNet In 1 Hour (Presentation)
Training with TPUs (Presentation)
Coffee Break (Break)
KFAC and Natural Gradients (Presentation)
Neumann Optimizer (Presentation)
Evolutionary Strategies (Presentation)
Future Hardware Directions (Discussion Panel)
Learning Device Placement (Presentation)
Scaling and Sparsity (Presentation)
Small World Network Architectures (Presentation)
Scalable RL and AlphaGo (Presentation)
Scaling Deep Learning to 15 PetaFlops (Presentation)
Scalable Silicon Compute (Presentation)
Practical Scaling Techniques (Presentation)
Designing for Supercompute-Scale Deep Learning (Presentation)
Adaptive Memory Networks (Poster Session)
Supercomputers for Deep Learning (Poster Session)