The impact of differentiable programming: how ∂P is enabling new science in Julia
summary
Julia now exposes ∂P in Julia so full programs — including differential equation solvers and lookup logic — yield exact gradients. A trebuchet simulator computes distance in ~2 ms and returns gradients for wind/angle/weight for ~0.5 ms extra (finite differences would need six runs); simple gradient descent reaches ~30 cm in 10 iterations. Training a small neural controller gives microsecond inference and ~1,000× speedup, wide-range errors ≈1–2 m in under a minute on one core; other wins include 15,000× speedup for cloud microphysics surrogates.
AI-written summary · about
chapters (9)
- 0:00Introduction to Julia
- 1:05Derivatives and Calculus
- 4:42Differentiable programming
- 8:41The trebuchet problem
- 11:41Physics and derivatives
- 17:14Neural network integration
- 23:32Modeling complex systems
- 29:11Scientific machine learning
- 43:43Q&A and conclusion
related talks
54mGeneralized Physics-Informed Learning through Language-Wide Differentiable Programming
59mDifferentiable Programming in C++
36mAccelerating Differential Equations In R and Python
43mJulia: A Fresh Approach to Machine Learning
27mJulia (workshop)
30mProductivity and performance in one programming language - A tour of Julia
01h 27mIntroduction to Julia for Scientific Computing and Data Science
39mDifferentiable Functional Programming