Ask HN: Is AI/ML over-hyped in biology and earth science?

3 pointsjarenmf2 comments

I see a general trend in academia to include machine learning in every proposal and apply it to every problem. I failed yet to see the advantages outside prediction which is some cases are indeed very useful e.g. the protein folding but I don't see how far it can advance science. My question is, do you think ML in academia is currently over-hyped?

View on HN

I don't know those specific fields but when I was in grad school, I saw a lot of ML applied to biomedical signal processing. Firstly, I think there is a ton of good in ML, specifically the ability to pull out (and rely upon but not reveal) patterns that a human couldn't. But what I saw as a challenge was when causal mechanisms, or the underlying physics is ignored, and ML is used to find a correlation. For meaningful application of ML to a discipline like you name, i think there should be a clear physical mechanism that underlies whatever ML is being used for, not just "we trained a model and it predicted this". Otherwise, ML (and gradient descent trained DNNs in particular) are very effective tools.

I think machine learning is currently overhyped everywhere and I think many machine learnign reserachers will agree.

My concern is more that scientific results derived by machine learning can't be relied upon. With neural network classifiers in particular it's easy enough to get them to overfit to any dataset, even random noise, so I don't think we can really learn anything useful from the performance of a classifier on some dataset, other than the fact that the classifier performs well on the dataset.

Reply on news.ycombinator.com