I'm super excited by how close we're getting to releasing some work I've been doing at Stripe on a static compute graph + bayesian inference library for Scala. Depending on your background, you might think of it as aspiring to be "TensorFlow for small data" or "Stan on the JVM".
-
-
@avibryant does that mean you have AutoDiff on arbitrary Scala working!!? We’d still likely use both your thing and ScalaStan. We’d love to private beta (and hope you’ll consider EvilPlot for your Bayesian posterior and diagnostic reports!) -
1/ EvilPlot looks completely awesome and I am eager to get it integrated; and please email me to get into the beta, I would hugely appreciate your feedback.
-
2/ it's autodiff on arbitrary Scala that operates on a numeric type which represents an unknown double and does not allow you to directly inspect its value (so you can't, say, use it to index into an array - just do math).
-
3/ whatever your Scala code is that manipulates these, ultimately the function it produces (and its gradient) gets compiled down to allocation-free generated JVM bytecode for Array[Double] => Array[Double]
End of conversation
New conversation -
-
-
How easy is it to add new inference algorithms? How easy is it to compose existing inference algorithms?
-
We've really given no thought to composing inference algorithms so I wouldn't expect that to be easy. In terms of adding new inference algorithms, I guess it would depend a lot; eg particle filtering, I think would be pretty hard to add; more variants of MCMC would be easy.
-
Ironically the starting point was a Scala implementation of https://dl.acm.org/citation.cfm?id=2804317 … which has a focus on composition and also particle filtering, but we've diverged greatly from that now.
End of conversation
New conversation -
-
-
I so want to see how much better/worse this works than Scalastan
-
Have you used Scalastan for anything?
-
It is developed at my current employer (which also released EvilPlot). We use it for real things, but a science startup is very different from Stripe
-
Oh cool. I’d be very interested to get feedback from you and/or your colleagues (even if it’s just examples showing why scalastan is better :) Send me an email if you want early access.
End of conversation
New conversation -
-
-
What are y’all using it for?
-
Time series forecasting.
-
BTW would be very curious to hear your feedback (either now or once it's released) - your book was a very useful resource while we were working on it.
-
thanks for the kind words =) I'm working on a side-bayesian project now. So I'd be interested in seeing it now, but probably won't have much time for a few weeks to give feedback.
End of conversation
New conversation -
-
-
Is it reasonable to use it from another JVM language, not Scala?
-
Interesting question. It hasn't been designed with that in mind, and the high level modeling API is very Scala-specific, but I suspect someone could build a shim for the lower levels.
End of conversation
New conversation -
-
-
me me me
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
I’m interested.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.