Soumith Chintala at AI Frontiers: A Dynamic View of the Deep Learning World
8 likes1,469 views
The document provides an overview of Torch, an interactive scientific computing framework similar to MATLAB and Python with strong GPU support for neural networks. It discusses the current state of AI, including applications in classification, segmentation, and reinforcement learning, as well as future research directions like dynamic neural networks and online learning. The core philosophy emphasizes flexible, imperative programming with a focus on community-driven development and easy integration with various environments.
Soumith Chintala at AI Frontiers: A Dynamic View of the Deep Learning World
- 1. Soumith Chintala Facebook AI Research A dynamic view of the deep learning world
- 2. Overview • What is Torch? • The Community • Today’s AI • What Next?
- 3. What is ? • Interactive Scientific computing framework
- 4. What is ? • Interactive Scientific computing framework
- 5. What is ? • Similar to Matlab / Python+Numpy
- 6. What is ? • Easy integration into and from C • Example: using CuDNN functions
- 7. What is ? • Strong GPU support
- 8. Neural Networks • nn: neural networks made easy • building blocks of differentiable modules
- 9. autograd by • Write imperative programs • Backprop defined for every operation in the language
- 10. Distributed Learning • Multi-GPU and Multi-Node • distlearn by • MPI-like model • scales to a large amount of nodes
- 11. Core Philosophy • Interactive computing • No compilation time • Imperative programming • Easy to understand, think and debug • Minimal abstraction • Thinking linearly • Maximal Flexibility • No constraints on interfaces or classes
- 12. Community
- 13. Community
- 14. Community
- 15. Community
- 16. Community
- 17. Community
- 18. Community
- 19. Community
- 20. Today’s AI • Reasonably solved to be used in production
- 21. Today’s AI • Classification and Detection • in images, videos, volumetric data, sparse datasets DenseCap: Johnson et. al. https://github.com/jcjohnson/densecap
- 22. Today’s AI • Segmentation DeepMask: Pinhero et. al.
- 23. Today’s AI • Text Classification (sentiment analysis etc.) • Text Embeddings • Graph embeddings • Machine Translation • Ads ranking
- 24. Today’s AI static datasets + static model structure
- 25. Today’s AI static datasets + static model structure model does not change over training time
- 26. Today’s AI static datasets + static model structure model does not change over training time offline learning
- 27. Active and Future AI Research • Agents training in different and new environments
- 28. Active and Future AI Research • RL Agents in different and new environments • Example: OpenAI Universe
- 29. Active and Future AI Research • RL Agents in different and new environments • Example: OpenAI Universe • Online Learning • Example: self-driving cars
- 30. Active and Future AI Research • RL Agents in different and new environments • Example: OpenAI Universe • Online Learning • Example: self-driving cars • Dynamic Neural Networks • self-adding new memory or layers • changing evaluation path based on inputs
- 31. Active and Future AI Research • RL Agents in different and new environments • Example: OpenAI Universe • Online Learning • Example: self-driving cars • Dynamic Neural Networks • self-adding new memory or layers • changing evaluation path based on inputs • Structured Prediction • Viterbi style decoders
- 32. A Next-Gen Framework • Interop with environments • Like Universe, VizDoom, etc. • Easy data loaders • Rich scientific ecosystem (such as SciPy) • Neural Networks are not the only methods to be used • Dynamic Neural Networks • with no specific or static structure • Minimal Abstractions • better debugging and low-level programming in the hands of researcher • Graph Compilers • To speed up structured prediction and fuse operations • Numba, XLA, Theano
- 33. Let’s share! Torch Caffe TensorFlow Theano MXNet, etc. Eigen TH CuDNN MKL Numba XLA NNVM MPI NCCL