Hacker News Reader

Comment by sgstfevsh | Hacker News Reader

For folks that want to start making sense of what's being said I recommend going through d2l.ai, fast.ai, and deeplearningbook.org.

Most of contemporary AI is really just a combination of probability theory and some upper division math classes in an executable form. None of it is magic and the more people that know the vocabulary the less likely people are to buy into the hype.

If you want a high level overview of all this then Melanie Mitchell has a good book as well: https://melaniemitchell.me/aibook/. She does a really good job of putting everything into the right context and dispelling the marketing hype about the coming singularity and human obsolescence. In one of the chapters she covers deep reinforcement learning and it's one of the best high level explanations I've come across yet.

View on HN

> Most of contemporary AI is really just a combination of probability theory and some upper division math classes in an executable form. None of it is magic and the more people that know the vocabulary the less likely people are to buy into the hype

For those of us who think probability theory captures a large fraction of what intelligence does, this reads a lot like "don't believe the hype behind nuclear weapons! It's really just physics and some upper division chemistry in an explosive form. None of it is magic..."

Maybe this argument will convince people who think that the human brain is magic? But for those of us who think that the human brain is not magic, but proof that general intelligence is actually computable and only needs 20 watts... If probability theory and upper division math in executable form is all it takes to be dangerous, that wouldn't be surprising.

It only takes physics and upper division chemistry in explosive form to level a city, after all. Not magic.

What do you suppose is dangerous about linear algebra and basic multivariable calculus? I personally don't think there is any danger in more people knowing more linear algebra and calculus.

The nuclear bombs were an analogy, and the point it is trying to prove is not "danger", but "powerful".

I use the term in the sense of the common phrase "know enough to be dangerous".

While the sources you mention are all good, I don't think any of them actually have much to say about what the link is about, which is bayesian deep learning.

That's correct but the people that are interested need a place to start and those sources are as good a place as any. If you have good learning material for Bayesian approaches then you should post them because whoever is reading the comments might take note.

I'm setting up a reading group at work.

Would this book be a good one for that?

Any other recommendations?

The Melanie Mitchell's book is mostly a high level overview. It's a really good introduction to the field but you will not become a good practitioner by just reading the book. If you want the mathematical details then d2l.ai so far seems pretty good but they use MXNet for the code. fast.ai uses PyTorch and in the later lessons introduces Swift4Tensorflow which bakes differentiable programming into the compiler. PyTorch is more popular so more jobs are likely to have code bases that use PyTorch.

So I'd say it depends on your use cases. If you want high level understanding then go with Melanie Mitchell's book. If you want more technical details then d2l.ai or fast.ai are good choices.

There's also "Deep Reinforcement Learning Hands-On" by Maxim Lapan. That one is also pretty good but I'm not far enough along to give a review one way or another.

Thank you - good tips.

> ... just a combination of probability theory and some upper division math classes ...

wow thanks for opening my eyes.

I guess I should just get together with my undergrad friends, implement the next DeepMind competitor, and collect our $400m from Google.

Well... yeah, kinda. Depth of domain knowledge and good engineering practice remains more important than individual genius in getting these things to actually work.

Agree - also business and people management.

The funds must flow, the people must stay, and work. Together.

Vision (the right vision) is a huge differentiator and ingredient x. I think you (I) can pull together a team with knowledge, practice, management and people structures, and make it work. The folks who do epoch making stuff have the amazing vision thing as well.

Reply on news.ycombinator.com