Hacker News Reader

Comment by etaioinshrdlu | Hacker News Reader

I think with good simulation tools, optimization processes, maybe even deep learning, we could make it a lot less of black magic. Make it boring and approachable instead.

It kind of reminds me of chemistry. You have physical laws that are fairly simple but the interactions are hard to describe without a bunch of computation.

View on HN

How does slathering deep learning onto something make it less of a black magic?

That's a good point, but given the search space is a set of voxels, you can probably brute force a yuge amount of possibilities and fit some sort of paramterized curve to them.

That's the sort of talk that gets the villagers sharpening their pitchforks.

Well deep learning is just a good way to make universal function approximators. So that's not magic.

I don't how know why neural nets get such a bad rap, they work well for many problems.

What’s the point of applying deep learning to something for which we have an exact mathematical theory?

Answer: because the exact mathematical theory can't give you models, just the performance of a given model.

Calculating EM equations exactly is also a GIGANTIC PAIN IN THE BNC connector.

It's much sexier if you add some deep learning

The analysis part of engineering is, thanks to modern computers, practically "done". However, we are still not able to do synthesis well. We also have exact mathematical models for that, but it usually amounts to NP-hard problems (think integer optimization).

I don't know about this specific case, but often the exact models are so complex that you can't optimize them analytically, so you resort to a global optimization best effort, where the results are not guaranteed to be any good - so you might as well try something DL-based too, to see how it would perform.

Classic traveling salesman problem cast into a large number of analytically incompatible and often hard to quantify dimensions.

I'm not sure how useful DL is for optimization, but other techniques such as Genetic Algorithms or Simulated Annealing would make more sense, I think.

Indeed. The evolved antenna page has an example. https://en.wikipedia.org/wiki/Evolved_antenna (previously here: https://news.ycombinator.com/item?id=18449434).

GAs are even more “black magic” than DL

I saw a NASA Tech Brief from 25 years ago where they did that for a small multi-band satellite antenna. The result looked like a pin cushion someone sat on.

It's been done: https://www.researchgate.net/publication/228909002_Automated...

>maybe even deep learning, we could make it a lot less of black magic

these two things are antithetical

Reply on news.ycombinator.com