Data Science Programming December 2020 Newsletter
This month is a special edition dedicated to JAX! It's a Python package built by some friends I made while they were at Harvard's Intelligent and Probabilistic Systems lab, and I was still in grad school.
I've been a fan of JAX ever since I started seriously developing array programs that required the use of automatic differentiation. What's up with JAX, you might ask? It's a library that brings automatic differentiation and many other composable program transformations to the NumPy API.
Why is automatic differentiation significant? The reason is that the ability to calculate the derivative of a function, w.r.t. one or more of its arguments, is essential to many computation realms. For example, we can use gradient-based optimization to train small and large models to do maximum likelihood or maximum a posteriori estimation of model parameters. Gradients are necessary for modern MCMC samplers, which leverage gradients to guide where to draw a new posterior sample next. Input design problems can also use gradient-based optimization, in which we either optimize or sample new inputs to achieve some output.
What JAX does is it takes a function that returns a scalar value and returns the derivative of that function's output w.r.t. the inputs. JAX accomplishes this by using the
grad function, which takes the function passed into it, and transforms it into another function that evaluates the gradient. Gradient transformations are one example of a broader class of program transformations, which take a program (e.g. a function implemented in NumPy code) and transforms it into another program (its derivative function). JAX houses other program transformations, including just-in-time compilation for speed-ups, loop-replacement functions, and more.
Here, I'm going to highlight a sampling of the JAX projects that have come up on my radar to showcase the diversity of numerical computation projects that you can build with it. Hopefully, it'll give you some encouragement to give JAX a try if you haven't already done so!
Neural network projects
Because differential programming is a broader thing than just neural networks, you can write neural networks and more using JAX. If you're not used to writing neural network models from scratch, not an issue: there are a few neural network API frontends that build on top of JAX's NumPy API, which implements PyTorch-like APIs.
flax: A neural network library focused on flexibility.
haiku: One developed by the fine folks at DeepMind, alongside their other JAX projects.
stax: JAX's internal experimental module for writing neural network models, which pairs well with its
neural-tangents: Research that I have been following, one that provides "infinitely wide" versions of classical neural networks. It extends the
The best part of these projects? You never have to leave the idiomatic NumPy API :).
Probabilistic programming projects
As someone who has dabbled in Bayesian statistical modelling, probabilistic programming is high on my watch list.
The first one I want to highlight is PyMC3. More specifically, Theano. One of our PyMC devs, Brandon Willard, had the foresight to see that we could rewrite Theano to compile to JAX, providing a modernized array computation backend to Theano's symbolic graph manipulation capabilities. It's in the works right now! Read more about it on a blog post written by the PyMC devs.
The second one I want to highlight is NumPyro, a JAX-backed version of the Pyro probabilistic programming language. A collection of Pyro enthusiasts built NumPyro; one of its most significant selling points is implementing the No-U-Turn Sampler (NUTS) in a performant fashion.
The third one I want to highlight is
mcx, a learning project built by Remi Louf, a software engineer in Paris. He has single-handedly implemented a probabilistic programming language leveraging JAX's idioms. I had the privilege of chatting with him about it and test-driving early versions of it.
Tutorials on JAX
Here are two tutorials on JAX that I have encountered, which helped me along the way.
Colin Raffel has a blog post on JAX that very much helped me understand how to use it. I highly recommend it!
Eric Jang has a blog post on meta-learning, with accompanying notebooks linked in the post, that show how to do meta-learning using JAX.
Beyond that, the JAX docs have a great tutorial to help get you up to speed.
From my collection
As I've experimented with JAX and used it in projects at work, here are things I've had a ton of fun building on top of JAX.
The first is
jax-unirep, done together with one of my interns Arkadij Kummer, in which we took a recurrent neural network developed by the Church Lab at Harvard Medical School and accelerated it over 100X using JAX, while also extending its API for ease of use. You can check out the pre-print we wrote as well.
The second is a tutorial on differential programming. This one is one I'm continually building out as I learn more about differential programming. There are a few rough edges in there post-rewrite, but I'm sharing this early in the spirit of working with an open garage door. In particular, I had a ton of fun walking through the math behind Dirichlet process Gaussian mixture model clustering.
Thank you for reading
I hope you enjoyed this JAX edition of the Data Science Programming Newsletter! Next month, we resume regular scheduled, ahem, programming :). If you've enjoyed this newsletter, please do share the link to the newsletter subscribe page with those whom you think might benefit from it.
As always, let me know on Twitter if you've enjoyed the newsletter, and I'm always open to hearing about the new things you've learned from it. Meanwhile, if you'd like to get early access to new content I make, I'd appreciate your support on Patreon!
Stay safe, stay indoors, and keep hacking!