Comment by minimaxir | Hacker News Reader

Comment by minimaxir | Hacker News Reader

> However, you don't need a degree at all (a few of our users are self-taught high school dropouts), and anyone who knows the material is welcome.

> To qualify for joining, you do an online deep learning quiz (here: https://www.sharpestminds.com/members/apply), followed by a technical interview.

> Once you join you get access to a job board with exclusive (i.e., not scraped) internship and full-time opportunities on it.

These three constraints don't reconcile with each other.

Yes, new ML/AI resources like TensorFlow and MOOCs have made AI more accessible, and that having a degree is no longer required to implement ML/AI. I agree it's unnecessary gatekeeping to require a degree to be able to play with ML/AI.

But what showy YouTube videos and Medium thought pieces don't teach is implementing ML/AI in practice to solve business problems. The stereotypical quiz + technical interview for the ability to join the service won't account for that.

When I was looking for jobs last year, 100% of the job openings for ML/AI (as opposed to Data Analyst/Data Scientist) required a Masters/PhD. In that case, I can't blame them, since there is a certain amount of experience and knowledge required to define problems and work up statistically sound solutions that can't be done by simply adding layers to a neural network or ensembling XGBoost models.

You're right, but this is the same catch-22 for getting started in any nontrivial field. How do I get experience if getting a job requires experience?

Having an MSc / PhD in the field doesn't resolve this. HR departments use grad degrees as first-pass filters, and thereby miss self-taught people who are genuinely competent.

We try to solve for this by easing people into jobs with internships and work terms first. The community is a key part of that since it supports them if they get stuck on an implementation problem. And of course we're incentivized to make sure members perform well in the internship phase, since we make money when they're hired full time.

I completely agree. Machine Learning is on the way to become a field like Web Development. There is a huge supply/demand gap that will only get wider.

With the caveat that holding a PhD is positively correlated with success in ML, and negatively correlated with success in web dev.

(I'm guessing.)

Having done both with varying success, I definitely agree that your brain needs to be wired differently for each.

In web dev, experimentation is cheap, so you make changes fast and see what happens. In many ML applications, trying stuff is expensive in time and/or money. So the best strategy is often to think hard about what could be going wrong, and make and test explicit hypotheses.

The difference definitely came as a surprise to me when I was making the transition.

Yeah, web dev seems to favor the bold, whereas DL or other highly brainy endeavors seem to favor the thorough / careful developer.

For a long time as a non-web dev, I reckoned that I was smarter than 99% of the web devs. Then a web-dev friend of mine made his first few million $, and I reached some humbling conclusions.

Sounds like I should start moving towards ML...

In this case there is no catch-22, you get experience by getting a Masters or a PhD. The truth is that there are enough people graduating with PhDs, and even more with Masters, that no one really needs to bother with people without a high school degree. The signal to noise ratio even for people with "just an undergraduate degree" is so high that I can't imagine even going to that level.

I always felt it was the other way around - using MOOCs or online tutorials makes applying machine learning to business problems in practice a whole lot easier. But in terms of inventing new algorithms, deeply understanding the theory and origin of things, and doing active ML/AI research, that still seems to be in the realm of requiring a PhD (i.e. the difference between a "data scientist" and an ML/AI researcher).

That's true: the reason I started looking into AI is to broaden my knowledge on how it can be used to solve problems in ways traditional ML/traditional modeling can't. I did have a strong statistical/data background in college beforehand which helped validate which approaches were good/bad/wrong, though.

The YouTube/Medium posts however advocate "Learn Machine Learning from scratch in 3 Months!" which is a problem.

How long does it take to 'learn machine learning'? I'm pretty sure I know some about it. I can certainly apply some basic algorithms and there's no shortage of info about lots of variations on the basics I do already know. So what am I missing? How much would I have to learn to be able to claim, per your unstated standard, that I've 'learned machine learning'?

Reply on news.ycombinator.com