To solve PDEs, you can learn at a fixed resolution (e.g. with U-net) and interpolate. But error cannot be arbitrarily small this way. Our neural operators are guaranteed to approximate arbitrarily well at any resolution since they learn in function space https://
arxiv.org/abs/2108.08481
RESEARCH
-
Neural Operators for PDEs: Resolution-Independent Function Space Learning
By
–
-
FastAI Students Explain Diffusion Model Paper
By
–
Another great diffusion model paper explainer from our wonderful @fastdotai students!
-
NVIDIA Internship: Computer Vision and AI for Autonomous Driving
By
–
Come intern with us at @nvidia We collaborate with @ALVAREZ_JOSEM team to build full-stack computer vision and AI models with applications in autonomous driving,e.g https://
arxiv.org/abs/2204.12451 @ZhidingYu @voidrank -
Zero-shot Super-resolution Tests for Neural Operators Benchmark
By
–
But it is crucial to add zero-shot super-resolution tests. Neural operators can output at any resolution since they learn mappings between function spaces, but U-Nets cannot. Can u please add these tests to the benchmark? @rejuvyesh @jo_brandstetter
-
Google and Lund University’s Optimus Learns Optimization Architecture
By
–
Google & Lund U’s Optimus Learned Optimization Architecture Efficiently Captures Complex Dependencies https://
syncedreview.com/2022/12/07/goo
gle-lund-us-optimus-learned-optimization-architecture-efficiently-captures-complex-dependencies/
… -
Recent papers on LLMs for formal mathematics and theorem proving
By
–
We had two papers on this topic recently, one using a MCTS approach on formal systems (such as Lean or Metamath): https://
arxiv.org/abs/2205.11491 and one where we combine LLMs with formal provers to verify the generations: https://
arxiv.org/abs/2210.12283 -
PDEArena Benchmarking and FNO Architectural Improvements in PyTorch
By
–
I agree. Great to see PDEArena as a convenient benchmarking environment for PDE in @pytorchlightnin Also great to see further architectural improvements to FNO by integrating it within U-Net like structure. Would be great if PINO can also be added https://
arxiv.org/abs/2111.03794 -
Early ML Extraction Story: Manager Unhappy with 95% Accuracy
By
–
Story from the early days of ML in information extraction:
Manager: So how good is your ML-based extractor?
Engineer: 95% accurate.
Manager: How dare you? Don't come back until you've debugged it. -
CICERO by Meta AI Team Reddit AMA Tomorrow at 10am PT
By
–
Don’t miss it, tomorrow the team will be answering your questions on Reddit — our #CICERObyMetaAI AMA starts at 10am PT. Ask us anything!
