MIT and IBM Find Clever AI Ways Around Brute-Force Math | Hacker News Reader

MIT and IBM Find Clever AI Ways Around Brute-Force Math

spectrum.ieee.org

78 pointsBrajeshwar9 comments

Some comments on this publication: Their approach is to train a learned downsampler of the boundary and initial conditions, and pass it through a learned coarse-grained "simulator", but the output is only a specified target property, not the full PDE solution. The claim is a decrease in training data needed and faster execution.

The limitation is that you need to have a target property of interest which is not always the case when trying to simulate something.

The paper as it is written is very low quality. They basically ran no real benchmarks and ignored all the other work on this topic. They used a plain NN to act as a baseline and their justification for this was that if you used any of the other published surrogate models to predict a target property you would end up with a NN. This is not a good faith comparison at all. Besides that, they also only ran very tiny experiments which is not where the need for surrogate models lies. And even the layout of the paper is very chaotic.

Unrelated question: Is Nature considered a low quality journal for ML papers?

> Unrelated question: Is Nature considered a low quality journal for ML papers?

It isn’t published in Nature, it is published in Nature Machine Intelligence.

Nature has sprouted all these offshoots which are less prestigious than the original (although trying to benefit from its good name). Their prestige/etc varies depending on the specific offshoot in question. But often they aren’t one of the top journals in that specific field

Some or most of the talent doesn't have affiliation or a journal application fee.

Affiliation with a university or research institution was necessary for how many of the top journal articles in a given year?

Otherwise, people publish docs, whitepapers, open specs, and code with tests; and it's great regardless of journal.

JOSS publishes their cost structure. Figshare and Zenodo grant DOIs for git revisions. ORCID is optional and precedes W3C DIDs.

MIT has an atrocious ML department, all the papers I've read from them have been third-rate. Mid-tier colleges in Asia have better researched papers.

I must have missed something since the title of the linked IEEE Spectrum article calls out MIT and IBM yet the text of the article says nothing about either. The only source mentioned is the lead researcher who's from Georgia Inst of Tech.

If you look at the paper[1] the authors are 1 from Georgia tech, 1 from IBM "Watson", one from IBM Watson and MIT and 1 from MIT.

[1] https://arxiv.org/pdf/2111.05841v3.pdf

At this point, publishing a paper is just another college level assignment... for MIT students. Everyone there wants that named accomplishment, and students are just trying to hack their way into it.

https://arxiv.org/pdf/2111.05841.pdf

> ... Here we present a “physics enhanced deep-surrogate” (“PEDS”) approach towards developing fast surrogate models for complex physical systems, which is described by PDEs. Specifically, a combination of a low fidelity, explainable physics simulator and a neural network generator is proposed, which is trained end-to-end to globally match the output of an expensive high-fidelity numerical solver. Experiments on three exemplar test cases, diffusion, reaction–diffusion, and electromagnetic scattering models, show that a PEDS surrogate can be up to 3× more accurate than an ensemble of feedforward neural networks with limited data (≈ 103 training points), and reduces the training data need by at least a factor of 100 to achieve a target error of 5%....

repo link: https://github.com/payel79/PEDS/tree/publish

Reply on news.ycombinator.com