Neural Compiler Translates Programs into Differentiable PyTorch Modules
A new innovation known as the Neural Compiler has been introduced by researchers, designed to convert programs that utilize a Scheme-like expression language into optimized, differentiable components compatible with PyTorch. This advancement ensures the retention of floating-point precision from the original programs while facilitating gradient computations via autograd. In hybrid modeling scenarios, the compiled components accurately depict established physical laws while allowing for adaptations in response to uncertainties. The researchers validated this compiler across six diverse domains, including physics, dynamics, and heat transfer, marking a significant leap in merging traditional physics with data-driven approaches in scientific machine learning.
Key facts
- The Neural Compiler translates programs into differentiable PyTorch modules.
- Modules match source program to floating-point precision.
- Gradients are provided through autograd.
- Evaluated on six domains: Feynman equations, Lotka-Volterra, damped pendulum, heat equation, 3D vector mechanics, compositional generalization.
- Compiled modules match hand-coded PyTorch implementations.
- Addresses combining known physics with learned components.
- Automates process that previously required hand-written code.
Entities
Institutions
- arXiv