Deep Learning in Mathematica

Has anyone tried using the new deep learning features in Mathematica 11 yet?

I realize it’s probably not as popular as Python/Lua since it’s expensive as hell, but a lot of universities have student licenses (including mine), so some of you guys might be able to check it out for free :slight_smile: Looking at their APIs, it feels extremely nice, especially with the feedback it automatically returns. For example, here’s a small snippet of their CIFAR-10 example

Everything is clickable and kinda works like TensorBoard, and it works on GPU with zero setup (simply passing TargetDevice -> "GPU"). I don’t want this to sound like a promo, since I really dislike a lot of things about Mathematica (especially their pricing model and closeness), but holy shit some of the features I wish we had in Jupyter.

If there was one thing to highlight, I’d probably say interactivity with data, since in Jupyter/Python even with matplotlib and numpy we still have to write quite a bit of code for things to show up properly even when just looking at stuff (plot_imgs from utils is a good example of something that would be nice to have as a oneliner baked in). Looking at this makes me drool a little bit.

Mind the unconventional syntax, it basically calls model.predict and displays the predictions. It automatically shows the images as images since everything is symbolic and it knows how to display symbols.

Anyway, I just wanted to show this since it feels like a nice learning tool where one can experiment at a very high level in terms of code. Thought my biggest surprise was that GPU just works on Windows, without all the ridiculous setup one has to go through with Theano (https://github.com/Theano/Theano/issues/5348).

2 Likes