Looking for a crash course in deep learning

February 26, 2018 12:29 PM

I am applying for a product management job at a deep learning startup. I have a strong technology background but I am not a programmer or mathematician. I know generally what deep learning does, but not how it works, or how a programmer would evaluate or integrate a deep learning system into their products. I need to begin learning all of that.

What should I read to get myself up to speed on deep learning technology and the software platforms that support it? I'm hoping to understand the technology and terminology at a client level (i.e. be able to understand in general a deep learning package's API), know something about who the players are, and understand what check-box features a deep learning package could have.

(Okay, I'll it admit it feels sort of dumb to be applying for a job in a field that I'm saying I know nothing about. Suffice it to say that I have other skills and background that give me a shot. I'd like to maximize my chances by learning as much as I can, in case I do score an interview.)

Thanks.

posted by Winnie the Proust to Computers & Internet (5 answers total)

The Deep Learning book might be your best bet in terms of skimming theory. Andrew Ng has a new Coursera venture, but it's a very slow means of absorbing this information and I didn't find his other Coursera course to be much good (though that's a minority opinion).

Tensorflow does have a Getting Started for ML Beginners tutorial, but it's honestly unlikely to teach you anything. (I also don't know that Estimators are really representative of how most people are using Tensorflow at this point. They're new and the documentation is less than great.)

There's part of me that feels like you've asked the wrong question, but I'm having a very hard time articulating why. I think it's that the big neural network packages/frameworks aren't all that different in terms of features (and, besides, they've surely picked one already). There is a hole around integration with Apache Spark that will be filled sooner or later (there are various options, but I don't think people have coalesced around anything yet). Amazon has gotten behind MxNet, which is interesting, as it certainly feels like Tensorflow is rapidly becoming dominant.
posted by hoyland at 6:37 PM on February 26, 2018 [1 favorite]

I'm going to flip hoyland's answers; next person can break the tie! I think that even with a strong math and programming background the Goodfellow et al Deep Learning book would be tough sledding if you haven't been exposed to a lot of it already. I think that Ng's course is probably the most efficient way to get up to speed, but even that may be too much theory and not enough practice for what you want right now.

If by "deep learning technology and the software platforms that support it" you mean things like Tensorflow vs PyTorch vs MXNet, then I think that hoyland is correct that they've probably picked one already; also, the attributes that distinguish them are things that you're not going to understand well unless you get your hands dirty with something like the Ng course.

Everyone learns differently, but I think your best bet is to start by seeing how well you can digest the Ng course; it will at least give you enough background to understand the fundamental ideas and terms of the field.
posted by dfan at 7:11 PM on February 26, 2018 [1 favorite]

I'm a programmer but not a math person. I found this video tutorial by Martin Görner from Google to be really terrific. It is definitely technical, and you will probably need to use supplementary sources to fully follow it, but I got a lot out of it. Deep learning libraries are much lower level than other APIs you might be used to (as compared to say, databases). Even with toy problems I had to get closer to the metal than I would with the kind of web applications I typically build.
posted by nev at 10:10 AM on February 27, 2018 [2 favorites]

To be clear, I meant "skim the starts of chapters of the Deep Learning book" with a view to figuring out what to Google. There are decent tutorials out there for a wide range of topics, but nowhere I know of that does a great job of giving a good high-level overview.
posted by hoyland at 6:17 PM on February 27, 2018 [1 favorite]

Response by poster: Thanks for these pointers. It looks like there are some good starting points among them.
posted by Winnie the Proust at 7:46 AM on March 2, 2018