# Rust-ML WG Meeting 00037

# Meeting Info

Date: 2022-09-13

Start time: [1100ET](http://timesched.pocoo.org/?date=2022-09-13&tz=utc!,gb:london,de:berlin,us:new-york-city:ny,us:austin:tx,us:seattle:wa,cn:beijing&range=1140,1260)

Link: https://fau.zoom.us/j/68484823002?pwd=ODN4SytRZktzSURJOEptQUJnbWdEUT09

# Participants

Lorenz, Manuel

# Notes by Chris

I'm afraid I'm probably not going to be able to make today's meeting after all, or will otherwise be pretty late. Here are my updates:
- I've been reviewing a JOSS submission by Rémi for JOSS (which uses `linfa`, hooray!), and am nearly finished on that end
- I started an update to the Rust-ML Book at rust-ml.github.io/book, which updates everything to use linfa 0.6
- https://github.com/rust-ml/book/tree/linfa-0.6
- I'd like feedback from Lorenz on the possibility of removing mdbook-scientific from the repository, since it's not currently being used for anything and makes building kind of complicated. In addition, `mdbook` now has a pretty good plugin for MathJax, and plotters is being updated regularly enough that I think the gnuplot plugin for auto-generating plots isn't as useful as it once was vs. just rendering plotters-generated images or .svg files or something
- The current content is basically just KMeans and DBSCAN clustering algorithms--up next, I think that doing a simple linear regression would be be nice, since that's a pretty vanilla task.
- I've been keeping an eye on the `dfdx` repository, which seems to be making a lot of progress. No GPU yet, but the author has basic convolution 2D layers working, which is pretty neat. I'm hoping to experiment with doing CIFAR-10 classification with it, since historically a number of Rust-based NN libraries have done great on MNIST with fully-connected layers, then stalled out when it comes to implementing convolutional layers on RGB imagery. https://github.com/coreylowman/dfdx

# Notes by Yuhan

Couldn't make it today because of work, so I'll just give some updates.

- Did a quick review of the JOSS submission. Pretty much done on that end, and I'm not sure if I'll review it anymore.
- Waiting for progress on these active PRs in linfa:
https://github.com/rust-ml/linfa/pull/236
https://github.com/rust-ml/linfa/pull/242
https://github.com/rust-ml/linfa/pull/238
- https://github.com/rust-ml/linfa-linalg/pull/11 still fails for some problematic matrices, including the linfa PCA tests. I need some help with that.

# Meeting Notes

- we missed the starting time and the Zoom link didn't work \o/
- EuroFest beginning of October
- Manuel is attending
- Lorenz still considering attending, but no agenda public yet
- connecting to the Berlin Rust scene? Perhaps some interest in ML
- `mdbook-scientific` can be of course removed and was rather a proof of concept, it is much simpler to use MathJax
- JOSS issue + PRs @Lorenz will look into (was on vacation last week)
- we took a look at dfdx
- combines a lot of things at once (perhaps too much?)
- Manuel: const generics really interesting
- Lorenz: einops library written in Rust with const generics would be a dream
- enzyme progress
- Manuel fixed million bugs working on his thesis from clang frontend
- learning LLVM is kind of hard, but https://lowlevelbits.org/how-to-learn-compilers-llvm-edition/ good starting point
- Lorenz: to reduce compilation time, requires marking crates as autodiff (similar to proc-macro)
- Manuel: this may not be necessary, but may add a lot of complexity to the implementation
- stable MIR would allow implementing enzyme for all code generation backends
- Lorenz will start working on enzyme again
- Manuel gives in his thesis and has time again from beginning of October for two months