Lush – Lisp-like language for machine learning designed by Yann LeCun
lush.sourceforge.netIf I wanted to do ML but with lispy syntax I'd probably use Hy, which is just a python re-skin. There is plenty of literature out there about ML in Hy (although it's not hard to translate from python too).
For example: https://github.com/FPTensorFlow/hy-tensorflow-tutorials
I prefer https://github.com/gilch/hissp, where Hy has to use shims to pretend statements are expressions, Hissp just targets the expression subset in the first place. (though as you mentioned, hy has a lot of literature and support around it, where as you're going to have to find your own way around hissp)
You don't have to go it alone. Hissp does have documentation (which is pretty thorough) if you want to try, but there's also a community chat on Gitter/Matrix: https://gitter.im/hissp-lang/community
If you know how to do it in Python (or Hy, probably) the community chat can help show you how to translate it to Lissp.
The Lissp REPL also shows the Python compilation before evaluating it, so if you're already familiar with Python, that can help you get a handle on Hissp.
latest update appears from 2009? Including instructions on how to fetch from cvs? Makes me a bit nostalgic, but not sure if this is relevant anymore
Last update was Feb 2022: https://sourceforge.net/p/lush/code/HEAD/tree/
by Leon Bottou (one of the original authors).
Yann Lecun posted on twitter yersterday https://twitter.com/ylecun/status/1628386056641847296?cxt=HH...
>Hotter take: ML would have advanced faster if another front-end language had been available and widely adopted instead of Python. One that is interactive yet fast & compilable, multithreaded (no GIL), isn't bloated, doesn't care about white spaces,... E.g. Julia or some Lisp.
Yes, that’s the news item that probably promoted the submission.
Yann’s later chaser to the above take (https://twitter.com/ylecun/status/1628390522892025860) —
> Even hotter take: the fact that ML and Computer Vision researchers were largely using Matlab held back progress for years, mostly because implementing something like a ConvNet in 2005 Matlab would have been a total nightmare.
And there are arguments back and forth about that. It’s funny, though, because a lot of the leading 1990s-2000s ML work was default-Matlab and people mostly loved it. Kevin Murphy used to advise Mathworks on bayes net implementation, and a lot of researchers from that time really liked the “numerical shell” that the Matlab REPL offered.
But of course one can see Yann’s point that the lack of clean capability for abstractions (I think struct’s came in the later 90s, lambda’s in the mid 2000s?) and licensing costs did hold the field back.
I did the original machine learning coursera course from Andrew ng, the homeworks in matlab were a horrible experience, seared into my memory till this day.
fwiw, I'm not a researcher though.
Sorry, I did not check the repo itself, I don't think I even have a cvs client anymore. On the linked page there's a section prominently placed, titled "Latest News". If you check there, the latest of such news is titled "A first Lush 2.0 beta is out!" with a date of 2009-10-19.
IIRC there is a zip for the beta 2009 release.