It’s a year later and I’m back: this time to the West Coast for the next edition of JuliaCon. The latest edition promised more talks, workshops, and wider community attendance. As usual, the conference began with a day dedicated to workshops.
The first workshop was on the DifferentialEquations ecosystem, championed by Chris Rackauckas, who spoke about his vision to make the ecosystem the scientist’s one stop shop for simulations. It certainly looked to be shaping up into a robust ecosystem. Incidentally, Chris also has an army of GSoC students working on improvements to many different little pieces. The next workshop I attended was the one on
Optim.jl. Statistical learning problems are often minimization problems where the objective function to minimize is your error function. The Optim ecosystem provides a rich variety of techniques to solve this problem, with a rich selection of options and callbacks for the user to understand why his problem hasn’t converged.
The last workshop was on machine learning in Julia. It started explaining concepts in machine learning from the ground up and then coding them up in pure Julia. Then we went into neural networks and deep learning before writing a recurrent network in pure Julia.
The day in retrospect turned out to be quite heavy, but I had to ran back home to start preparing for my talks.
The first day’s keynote presented by Fernando Perez, a research scientist at Lawrence Berkeley National Labs (LBNL), was on using online software and tools such as Binder that can render Jupyter notebooks, and how this would help researchers share their plots and results in the form of a notebook. Stefan’s Pkg3 talk laid out the problems with the old package manager and then articulated how Pkg3 would solve some of those problems. At this point, I decided to go brush up for my joint talk with my colleague Simon Byrne on Miletus, a Julia package used to model financial contracts and then to value them using a suite of algorithms. In the afternoon, my colleague Jameson attempted to break down the inner workings of the Julia compiler and some future directions he might want to take it, after which Tim Besard spoke about his work on native GPU code-generation, where he intercepts LLVM IR to generate PTX code that runs directly on the GPU.
It was nice watching my colleague Mike Innes talk about Flux, which is a new Julia package for machine learning that aims to treats models as functions with tunable parameters. Through use of macros and functional programming paradigms, Flux’s ease of use would allow developers to design complex neural network architectures, and step through each layer like anyone would a normal Julia function. While Flux is meant to act as a specification library, a package like
KNet.jl aims to act like a proper computational backend. Deniz Yuret explained how
Knet.jl uses dynamic computational graphs and uses automatic differentiation to get the gradients of any Julia function. The package uses several high level language features which the author claims is missing from other frameworks that work with static computational graphs.
The second day started with Mykel Kochenderfer’s work with the FAA and Lincoln Labs on collision avoidance systems. He went into detail about the complex decision processes that he has to work with under uncertainty. This followed by Jeff’s review of the type system overhaul. This, for me, was one of the best presentations at the conference, for how Jeff broke down fairly complex ideas into a simple ones for general consumption. I also got a chance to listen for the first time to a talk on probabilistic programming, which was delivered by Kai Xu. The next talk was about multidimensional signal processing, especially focussing on the Shear Transform. I was delighted to know that the author, Hector Andrade Loarca, uses
ArrayFire.jl for his research, and that it helps him greatly.
The Celeste keynote was the highlight of the day, with my colleague Keno getting into the details of the computation. He spoke about how he had to make several improvements to the Julia compiler so as to make memory access patterns more uniform and hence more efficient, and how they even used Julia packages almost as is for the project. Not only was it remarkable that they crossed 1 Petaflop (which is probably the first for a high productivity language), they managed to do so without writing hand tuned Ninja code. This, for me, was the biggest achievement. After the keynote, I had to run off to prepare for my second talk at the conference:
Circuitscape, a landscape modelling tool that was used to calculate least resistance paths across large swathes of land. I went into detail about my vision for the package and how I will take it there, and hopefully earned a couple of collaborators.
At the end of the day, there was a poster session with GSoC students presenting their work over posters. They were all great. It was in particular nice to see an old intern of ours, Divyansh, do some great work on
LightGraphs.jl, where he parallelizes betweenness centrality and Dijsktra's algorithms with weighted graphs. Kenta Sato, who is a core contributor to
Bio.jl, was back as a GSoC student and was trying to get some parallelism into
The last day of the conference started Kathy Yelick from UC Berkeley talking about how the High Performance Computing (HPC) and Big Data worlds are sort of merging, and how researchers need to find ways to tackle larger amounts of scientific data. Then it was time for a lighter session with Jiahao talking about “how to take vector transposes seriously”. He went into the history of the famous issue #4774, much of which involved heated discussion on the different notions of the word “vector”. The afternoon started out with Stefan speaking briefly about the roadmap for Julia 1.0, assuring the community that 1.0 was still on despite the latest master being called 0.7-DEV. This was followed up by a very interesting session on why Julia is a great language for mathematical programming by Madeleine Udell, and focused on the package
Convex.jl, which infers and formulates convex optimization problems. Convex and JuMP are the two jewels on the crest of a robust optimization ecosystem in Julia.
The conference ended with Jeff Bezanson presenting a short talk on JuliaDB, the in-memory database that was part of the JuliaFin product from Julia Computing, and was recently open sourced. It supports both relational and SQL-like queries and uses
Dagger.jl under the hood. After Jeff finished, the question and answer session seemed to stretch for longer than the talk itself, indicating an unwillingness amongst the participants for the conference to end.
While it was certainly a hectic few days with meetings upon meetings with collaborators and plenty of information, I found that I could look back on the week with a sense of satisfaction. Looking back, I was pleasantly surprised as to how much the community had grown over the years: the breadth of topics, the size of the attendance, and most importantly, the warmth. I've a feeling this community will only get bigger, and every JuliaCon is proof of just that.