Papers of the day   All papers

Neural reparameterization improves structural optimization


Sep 28 2019 hardmaru

Neural networks for optimizing structural designs. Cool work!
7 replies, 590 likes

Sep 28 2019 Jascha

Neural reparameterization improves structural optimization! By parameterizing physical design in terms of the (constrained) output of a neural network, we propose stronger and more elegant bridges, skyscrapers, and cantilevers. With shoyer@ samgreydanus@
3 replies, 386 likes

Sep 28 2019 Stephan Hoyer

I'm happy to share a new paper, with @jaschasd and @samgreydanus: "Neural reparameterization improves structural optimization" We use neural nets to parameterize inputs of a finite elements method, and differentiable through the whole thing:
2 replies, 172 likes

Oct 15 2019 Frank A. Krueger

When you ask a neural network to optimize the structure of tall buildings, it invents trees.
5 replies, 124 likes

Dec 13 2019 Stephan Hoyer @ #NeurIPS

I'll be presenting this work with @samgreydanus and @jaschasd tomorrow morning (Friday Dec 13) at 11:30am at the NeurIPS Deep Inverse Problems workshop (
2 replies, 45 likes

Sep 28 2019 Benjamin Pope

Gothic architecture came from the genius of stripping away stone where it was in tension - what cybergothic architecture will we get from the dreams of neural networks?
2 replies, 13 likes

Sep 28 2019 Sam Greydanus

Our latest project is on arXiv! We explore the use of the “deep prior” for reparameterizing physics problems. It was a pleasure working with @jaschasd and @shoyer on this
1 replies, 9 likes

Sep 28 2019 Daisuke Okanohara

CNNs provides better parameterization for structural optimization (c.f. deep prior); inductive bias toward simpler and more effective structures. Backprop the error through the transformation at the optimum point using implicit differentiation.
0 replies, 7 likes

Sep 28 2019 NIDHAL SELMI - نضال السالمي

Neural Nets design stronger structures. Add 3D printing robots to this (already done in Amsterdam) and you get a picture of some future capabilities.
2 replies, 6 likes

Oct 15 2019 Stephan Hoyer

@praeclarum @jeremyphoward I agree! Optimizing complex simulations with auto-diff is a really powerful idea. We did some recent work in this area using neural networks to reparameterize structural optimization:
1 replies, 5 likes

Oct 15 2019 Suvy Bo

0 replies, 5 likes

Oct 24 2019 Shawn Presser

I'd love to see neural network videogame level design. Someone do it!
0 replies, 2 likes

Oct 16 2019 Alexander Kyte

So what you're saying is that tree houses are provably optimal engineering?
0 replies, 2 likes