Hacker News Reader

Comment by peterwoerner | Hacker News Reader

This kinda thing is prevalent in academia, it is probably totally useless and at best so incomplete as to not tell you anything.

There is a lot of research going back to the 1930s solving this problem--how do you estimate energy potential based on available data. The machine learning approach is really the variational approach but using very expensive (lots of parameter) functions instead of cleverly choosing your fitting function so that it has parameters which are physically meaningful and significantly fewer in number. The first helps with building a meaningful story of the underlying physics while the second means you need fewer data points.

View on HN

> cleverly choosing your fitting function

One benefit of ML is being able to replace clever subject matter experts (who may be expensive and hard to identify) with ML generalists.

Frequently, you still need a clever subject matter expert to figure out what form of input you need to insert into the neural network. E.G. someone has to figure out how to take crystal structure and put it in a neural network. Most obvious choices (e.g. voxel models) are intractable without huge amounts of computational resources which aren't generally available

Your data needs to be very cheap.

And you don't have to be that clever. For periodic function sines and cosines. For potential energy quadratics, polynomials and exponents do a great job.

One of the key components which makes neural networks generally inferior is that you need enough data to learn the governing equations (always true physics) as well as the constitutive law (part of the physics which applies only to your problem) as opposed to just the constitutive law.

Of course there are notable exeptions. Neural networks have been doing a great job of processing microscopy data and building 3d models from a series of 2d images.

And then the ML generalists (sometimes painfully) re learn the same lessons that the subject matter experts have learned over generations

While that’s the poster dream, I wonder whether it actually works in practice, rather than being a liability, as a sibling comment points out.

IME, imposing domain knowledge (usually directly into the model, or sometimes also in the training data distribution) is crucial to getting things working in an efficient and robust manner.

Deep NNs are already competitive with specialized pipelines for translation and speech.

The primary downside is that they require more data to train.

more importantly, thanks to the internet there is a huge amount of available data that is reasonably available. Many of the science problems I have dealt with have something like 11 data "points". And experiments may take a month to perform.

Sure, but that’s such a tiny corner among generic problems. To generalize from there seems unwarranted.

But is it a benefit or a liability? Reminds me a lot of the arguments in Susskind and Susskind's "The Future of the Professions."

yes, try fitting a function to less than pi/4 period of data...

I did this recently. We wanted to validate our thermal control model and we had ~20 heaters and 24 hours to do the test. Heaters couldn't cycle faster than about 4 hours due to thermal mass, so we had to fit our periodic function in less than a quarter of a period of cycle time. It works fine if you have a good initial guess to the period and you know your solution is periodic.

Reply on news.ycombinator.com