Keynote: Next-generation frameworks for Large-scale AI
summary
Claims that embedding structure beats scale: Tensorly layers and CP/tensor-train factorizations produce models an order of magnitude smaller with better accuracy and faster early-event detection; NVIDIA Isaac plus EGX/Jetson pipelines train multi-task robot policies at GPU scale in simulation and transfer to real robots (kitchen, BMW factory, four-legged robot) using hierarchical RL that handles slippery terrains; RAPIDS accelerates end-to-end GPU data pipelines; tensor-dropout and qtensor primitives improve robustness and hardware efficiency on Ampere-class GPUs.
AI-written summary · about
chapters (12)
- 0:00Introduction to next-gen AI
- 1:51Breaking down AI methods
- 3:18Limitations of current AI systems
- 5:47Next frontier in AI
- 7:58NVIDIA Isaac for robot learning
- 10:27Simulation to real-world robotics
- 13:39Incorporating robust priors
- 14:41Power of tensor methods
- 18:55Optimizing tensor models
- 20:00Bigger is not always better
- 21:47Tensors for deep learning
- 22:29NVIDIA platforms and tools
related talks
12mKeynote: What's New in AI
12mKeynote: AI Infrastructure- 12mKeynote: What's New in AI
28mNext-generation frameworks for Large-scale Machine Learning
32mKeynote: AI & the stuff built for AI - are they actually useful for data science?
40mFueling the AI Revolution with Gaming
12mKeynote: Build Accelerated LLM Applications with Open Source Tools from NVIDIA
26mKeynote: Artificial Innovation Accelerates Artificial Intelligence