If you can’t understand the proofs, then you’re taking what this educator says on faith. You may also have a less sophisticated idea of when to apply which methods. Your ability to evaluate new results / methods / etc may be compromised by your inability to evaluate them in a principled way, which may be facilitated by your understanding of their underpinnings.

On the other hand, it’s a rare work day when you derive a significantly new method / actually leverage the proofs / their underlying methodologies.

All in all, it’s like saying “you can do software engineering without understanding theory of computation”. You totally can, and can do it well, but you’ll have some blind spots that won’t be able to efficiently address / speak to your peers about.

There’s no one right answer. There’s the right answer for you.

Why bother learning? Download already implemented stuff and just run it.

Controversial opinion: A proof of an ML concept is not really ML. It’s (usually) mostly math. I don’t think knowing how to prove that gradient descent converges in certain conditions helps you apply gradient descent in any way whatsoever. It certainly doesn’t make you a better ML engineer, which is what most people learning ML are trying to do. Knowing that gradient descent converges in certain conditions is absolutely necessary, but the proof is not at all helpful to an ML engineer. You can gain an intuition for the concepts without the proofs, and I honestly don’t even think the proofs are very intuitive to begin with. I say all of this as someone who has been an ML student, an ML engineer, and an ML researcher in a professional capacity.

Here’s an anecdote. I took a convex optimization course during my M.S. program. It was a 10-week course, 30 total hours of lecture, and I think we only discussed maybe 10 theorems and / or methods related to convex optimization because we spent 3 hours proving the correctness of each one of them. Those 3 hours were spent doing relatively basic calculus and loads of algebra. ML engineers should obviously understand basic mathematics, but they certainly don’t spend a disproportionate amount of their time doing algebra. I would’ve much rather glossed over the proofs, sacrificing the math lessons to actually focus on convex optimization.

Unfortunately, there’s a disconnect—the teachers are often mathematicians (ML researchers; professors), and most of the students are engineers. And the teachers don’t do a very good job of appealing to their audience.

This discussion is actually deep. It is really a philosophical question: how low-level is enough just to use a piece of knowledge. Taking learning python as an example, if you just want to learn python for daily lightweight use, why bothering go deep to understand how built-in methods are implemented, how OS schedules jobs, how contexts are switched in CPU, how electrons move between silicons chips. Don’t get me wrong, all these are necessary and useful to become a python expert. But with limited time, you have to choose what to learn based on your need. IMO just learn the ones you really need now, and you will eventually get back to the proofs if you really need them in the future.

In my engineering school, a bunch of students don’t study math behind Machine learning but I believe spending some time in math makes you a better AI developer, but it depends on your purpose .