Hacker News Reader

Comment by probablybroken | Hacker News Reader

Wouldn't he also have to learn a lot of advanced mathematics? I'd imagine the bunch of graduates in this case would be math majors. Obviously I don't know the specific requirements of your organisation, but this re-skilling might take up to five years ( if the requirements were R&D focused... Maybe the situation is different if it's more about applying pre-existing sofware. )

View on HN

Does ML really need advanced maths? I think you're under-estimating how the fundamentals are pretty simple really. What we were previously lacking was raw processing power to get good results.

Yes, it does, especially if your going to put a model in production for anything critical ex: self driving cars, healthcare. Understanding the math is what allows you to properly understand what is going on, how to improve, debug, etc. A person who only knows how to use a library/framework will be limited.

Who said anything about him using it for health care or self driving cars?

The vast majority of applications for AI/ML can be done with libraries, just like the vast majority of programmers doesn't need to understand how to make a compiler or even how memory works. It just helps.

The vast majority of programmers do need to understand boolean algebra though. Likewise, you do need to understand matrix algebra and statistics to do any AI/ML beyond a color-by-numbers problem.

At some point you need to explain why something works.

Matrix algebra and statistics are taught in high school, or at least were when I was young.

A random 40 y/o programmer might be rusty on it, but it's not going to take very long to pick it all up again.

And if you've done any game programming even just for fun, you probably know quite a bit of that already.

There are plenty of ML problems that requires nothing more than being able to pick up something of the complexity of Bayes theorem. Yes, there are also ML problems that require far more. But you can't say much about the math requirements without knowing what kind of problems they'd be working on.

You would be surprised how the devil always emerges in the details when you move to real applications from academic exercises and datasets.

I definitely use a lot of statistics daily. I do not need to compute integrals by hand or construct formal proofs, but a lot of intuition is necessary.

Statistics is very tricky by the way. It is so easy to have incorrect assumptions and totally misapply the methods. Also another problem with ML in general is that if you make an error you very often do not notice it. The methods will usually work a bit worse but they will still kinda work.

Just yesterday I discovered a tricky feedback loop that corrupted my data for several weeks.

> Statistics is very tricky by the way.

I think this why HR hire computer engineers instead of statisticians when it comes to ML.

At least half of my colleagues have degrees in mathematics.

Reply on news.ycombinator.com