[atoav]

There is a certain value in understanding why something works and how you can either continously improve it or adjust a few dials when there is an exceptional situation.

Part of the fascination with ML is the (dangerous) myth that you don't have to wrap your head around a complicated problem anymore, instead the solution will just magically fall out on the other side of the blackbox if you just feed it enough data.

Understand ing the intricates of the problems you are dealing with however is a value in itself.

[mdp2021]

Yes, but (please allow me):

«Part of the fascination with ML is» solving the mystery behind the ability to automatically build functions and behind those functions.

Surely, both in practice and axiologically, understanding and deterministically solving have a great value. Also because of that, the fact that systems exist that can adapt into solutions, but contain a transparency problem ("yes, but why"), contains an immensely fascinating theoretical challenge, in the learning that may come from the attempt to understand the "grown, spawned" (as if a natural phenomenon) system.

The laziness is not necessary: there is a great deal of fascination in unveiling the mysteries in the blackbox.

Then of course, when you have a practical problem to solve (instead of that intellectual challenge and promise), pick your best solution. And surely it is sensible to call it dangerous to rely on something not properly understood, which may hide the potential faults ("yes, we found out it fails here, and it may be that we kind of assumed it "saw" shapes, while really it "sees" textures..."). In professional practice those "active" fascinations (understanding the spawned) may be luxury.

[wnkrshm]

While the methods are very interesting, I often wonder about assumptions about what can be modeled. We already know that it's not possible to correctly predict an arbitrary nonlinear system numerically (since that system could be chaotic).

It's one of the reasons why for specific problems, heuristics or statistics are way better than any attempt at nonlinear modeling / ML prediction (e.g. highly accurate climate models vs. struggling weather models).

[cm2187]

I even hear people pitching ML for applications where determinism and explainability aren’t optional, like regulatory and financial reporting for a financial institution.

[MikeDelta]

Indeed! Regulatory reporting is clearly defined, i.e., what needs to be reported and how. Yet I see groups trying to use ML to determine what to report, which makes me think they don't fully understand the topic they are working on.

[tomrod]

To be fair, explainability is still a hot topic of research, as well as discriminatory bias tradeoffs.