JuliaCon 2021: the main part starts now

[dmm]

juliacon.org/2021/ and live.juliacon.org/

The talk schedule: pretalx.com/juliacon2021/schedule/ (the default is UTC time, I see that I can switch to New York)

If you register (free), this gives access to discord, gathertown, and hackaton.

I'll keep notes in the comments.

Comments

[dmm]

8 hour days starting at 8:30 New York time, 3 days, then rumors of a hackaton on Saturday.

1.5 hour slot, 30 min break, 1.5 hour slot, 30 min break, 1.5 hour slot, 1 hour break, 1.5 hour slot - this seems to be the general scheme of things. Parallel sessions (but most of this should also be available for watching later on YouTube).

[dmm]

day 1

The state of JuMP

Agents.jl

Dictionaries.jl https://github.com/andyferris/Dictionaries.jl - this is an important topic for me

https://live.juliacon.org/talk/WRNAEN

https://pretalx.com/juliacon2021/talk/WRNAEN/

[dmm]

Discord was useful already (I was able to ask Andy whether he used it with Flux/Zygote, alas the answer is "no", so I'll be the first to try again ;-) ):

Michael Bukatin
—
Today at 9:49 AM
@ Andy Ferris Have you used it with Flux/Zygote (like using dictionaries instead of tensors in machine learning models)?
Andy Ferris
—
Today at 9:50 AM
No I haven't

Michael Bukatin
—
Today at 9:51 AM
I was trying to do that with Dict and with PeristentHashMap, and have not really succeeded so far; I think your package might work better in this sense...

[dmm]

I'll stay in green track mostly for the day, for keynotes and machine learning/differential programming talks.

[dmm]

Jan Vitek keynote is, among other things, about subtyping and their "rational reconstruction of Julia types" paper.

And their "World age in Julia" paper.

And "type stability".

***

Question answering: "jit compilers are machine learning programs" (the problem is that jit forgets things between runs, what it learned about libraries, about this particular program which is rerun and rerun".

[dmm]

Discord: Artem Pelenitsyn
—
Today at 11:27 AM
Please, checkout (very short!) Section 4.4 of the subtyping paper
https://fzn.fr/projects/lambdajulia/paper.pdf

Tldr 120 of 122 discrepancies were due to bug in Julia 0.6.2, it was fixed after we reported it. Two other cases were not decided upon. There are links two GitHub issues as well. (See appendix of the subtyping paper https://fzn.fr/projects/lambdajulia/issues.pdf)

[dmm]

Discussion of floating-point errors in the second keynote talk and on the discord channel is troubling, especially given its implications for machine learning programs.

[dmm]

The Trivia game (during the break) is remarkable!

[dmm]

ChainRules talk - a bit difficult to listen too, because it's too perfect (too polished, not "live enough"); but it's a great resource for mastering the ChainRules package (but one would need to stop it all the time, it's too fast).

And yes, ChainRules 1.0 is released!

[dmm]

Enzyme talk - interesting! (Derivatives for mutating arrays work, etc.)

Very, very interesting.

I should learn how to plug different AD systems into Flux (everyone probably knows this by the "word of the mouth", or smth).

[dmm]

Надо сказать, в таком режиме, когда параллельно с докладом идёт крайне интересное обсуждение этого доклада на дискорде, получается интересно, но очень уж насыщенно (и никакого шанса не утерять из памяти всякое).

В общем, я бы сказал, что это гораздо более интенсивно, чем конференции в реале... И ещё уровень участников разговора (рекордно высокий)...

[dmm]

Jonathan Laurent
—
Today at 3:08 PM
My research is basically on applying AlphaZero on the problems of automated theorem proving and program synthesis.

https://github.com/jonathan-laurent/AlphaZero.jl

[dmm]

I copied quite a bit of relevant discord discussion

[dmm]

The last interesting talk is not in green track: Pluto.jl notebooks are Web APIs!

[dmm]

Надо сказать, в таком режиме, когда параллельно с докладом идёт крайне интересное обсуждение этого доклада на дискорде, получается интересно, но очень уж насыщенно (и никакого шанса не утерять из памяти всякое).

В общем, я бы сказал, что это гораздо более интенсивно, чем конференции в реале... И ещё уровень участников разговора (рекордно высокий)...

[dmm]

Day 2.

I think I'll do the green track for the first three 1.5 hour session for machine learning and keynotes, but for the last 1.5 hour session (starting at 3pm New York time) I'll do the red track for visualizations.

[dmm]

Meanwhile, posters got numbers... There are 76 posters at the moment, and my poster is... number 13... hmmm...

[dmm]

SmartTensors: Unsupervised Machine Learning

Non-negative matrix and tensor factorizations are interesting! (E.g. if we want to invert our image mutiplication!)

There are multiple methods to do this. A very interesting lecture.

These methods extract features!

[dmm]

AutoML talk - a bit too technical, but it's impressive that he can do AutoML in real-time of the video - wow, wow...

https://github.com/IBM/AutoMLPipeline.jl

[dmm]

BoF on streaming live-coding sessions... I am not sure if that would be available later, because it's a BoF ("bird of feather")

[dmm]

State of Julia: opaque closures are actually like what we the Scheme people think closures should be!

Towards compiler plug-ins...

AD as a foundational capability: complimentary efforts.

[dmm]

Xiaoye (Sherry) Li keynote.

She shows "cluster trees" inside matrices... And she talks about methods for finding good bases...

She wants "even better support for sparse matrices"...

[dmm]

InvertibleNetworks.jl and normalizing flows

I like images showing at around 20 min of the video

Mentions "matrix-free linear operators" computed via propagators (all in all, this is an important talk)

[dmm]

Composable Bayesian Modeling with Soss.jl

Interesting (DAG, complex nodes)

[dmm]

Chaotic time series predictions with ReservoirComputing.jl

[dmm]

Generative Models with Latent Differential Equations in Julia

[dmm]

CompositionalNetworks.jl: a scaling glass-box neural network

[dmm]

I started to explore Gather Town during a break; found the poster rooms, my poster (#13) is in the room #1, no poster videos seem to be uploaded yet.

[dmm]

I also got a great advice of differentiating dictionary transforms on the "hackaton" channel of the conference Discord, namely to work on this: https://github.com/JuliaDiff/ChainRulesCore.jl/issues/186

Now to the red track and the last 1.5 hour segment:

Javis.jl - Julia Animations and Visualizations

Nice framework, good talk!

[dmm]

Now 3 lightning talks:

Julia and deploying complex graphical applications for laypeople

PGFPlotsX.jl - Plotting with LaTeX, directly from Julia

Towards an increased code-creativity harmony in Javis - nice layers! streaming to twitch!

[dmm]

A deep dive into MakieLayout

Impressive flexibility

[dmm]

Plan for Day 3:

Session 1:

red track: Symbolics.jl - fast and flexible symbolic programming

red track: Unleashing Algebraic Metaprogramming in Julia with Metatheory.jl

green track: JuliaSim: Machine Learning Accelerated Modeling and Simulation

Session 2:

keynotes and sponsor talks, starting with a keynote by Soumith Chintala !!!

Session 3: virtual poster session on Gather Town: https://juliacon.org/2021/posters/

Session 4: let's do the purple track

[dmm]

Symbolics.jl: https://arxiv.org/abs/2105.03949

[dmm]

Metatheory.jl: https://arxiv.org/abs/2102.07888 https://github.com/0x0f0f0f/Metatheory.jl

e-graph rewriting is tremendous; might be super-interesting for proof search/program synthesis/other AI problems as well (with some variations)

Very impressive!

[dmm]

JuliaSim - ambitious vision by Chris Rackauckas

(To be a commercial offering by Julia Computing; does not seem to be sustainable by the standard open-source volunteer model of development, says Chris)

[dmm]

Mentions continuous-time echo state networks: https://arxiv.org/abs/2010.04004

[dmm]

He is making very strong AI claims (that their AI is already generating ODE systems and such).

And he is an amazing speaker, but the feeling of "sales talk" is pretty strong... Even though most of those claims might just be exactly correct...

[dmm]

Soumith Chintala keynote at IBM AI Systems Day 2018 was great, super-inspiring.

This one is shaping to be a disappointment. He is just describing how huge this things now is, how much funding it pulls, and how "industrial" and "routine" this has all became (that's 15 min in, 1/3 into a talk). But now he is going to talk about open-source process, which might be interesting and useful...

User-centric vs technology-centric open source projects (Torch is user-centric).

and yes, when he moved to the open-source dynamics and remarks about that, it became quite interesting, even if low-key, but quite useful... He sheds a lot of light onto early period of PyTorch, which has been really great, and that's very interesting...

http://blog.ezyang.com/2021/01/pytorch-open-source-process/

[dmm]

Julia Computing is hiring on its new funding.

JuliaHub

JuliaSim

The whole single track session is concatenated as https://www.youtube.com/watch?v=6V6jk_OdH-w

[dmm]

Poster session is a mess in terms of tech glitches...

Still, I did talk to someone...

[dmm]

Applied Measure Theory for Probabilistic Modeling

MeasureTheory.jl

(by Chad Scherrer, the author of Soss.jl; a less interesting talk, too educational, and I am not a fan of the topic, I just wanted to hear what the author of Soss.jl would say)

He finds Dictionaries.jl useful!

[dmm]

FourierTools.jl | Working with the Frequency Space

Felix Wechsler, Rainer Heintzmann
07/30/2021, 3:30 PM — 3:40 PM EDT
Purple
Abstract

FourierTools.jl aims at simplifying work in Fourier/Frequency space without loosing efficiency. We provide several convenient wrappers to speed-up the common fft(fftshift(x)) pattern. This package also brings functionality to up and downsample signals through sinc interpolation. Furthermore, based on FFTs it provides shearing, rotation, convolution and (sub) pixel shift functions which can be applied to N-dimensional data efficiently.

Up Next:

IntervalLinearAlgebra.jl: Linear algebra done rigorously

Geometry of even 2D interval matrix mult is interesting

[dmm]

Global Sensitivity Analysis for SciML models in Julia

[dmm]

ZigZagBoomerang.jl - parallel inference and variable selection

A practically useful nice thing, to be used with Soss and others.