search

brainiak / brainiak-aperture

BrainIAK Aperture submission
Apache License 2.0
10 stars 10 forks source link

More MN prototyping examples? #24

Open mshvartsman opened 3 years ago

mshvartsman commented 3 years ago

Hi folks -

One of the cool things we've tried pushing with brainiak.matnormal is a prototyping tool people can use to try structured-covariance models more easily. With this in mind, it'd be awesome if someone wanted to kick the tires and see how easy it is to set up their favorite models in separable-structured-cov flavor. In the paper we had full examples with SRM and RSA, and in the notebook I've added matrix-normal factor analysis (basically slight extension of PPCA) just to show something new, but having someone else actually try it would be great. I think @jeremymanning and I discussed doing the same with TFA forever ago at SFN, and in my very naive understanding the existence of MatnormalRegression.calibrate makes that model a simple instance of IEM. Basically any (multi)linear gaussian regression or factor model should hopefully be not very much code to try.

jeremymanning commented 3 years ago

@mshvartsman this sounds interesting, but I'm not going to have time to do this over the next few days (so TFA as an example won't make it into this release, unless someone else takes the lead on it). Maybe I could entice one of my RAs to try it out next term, though. I also remember the conversation you're referring to, but I think it'd be helpful to do a zoom to help refresh those details. Let me know if you're interested!

mshvartsman commented 3 years ago

Yeah definitely not this release, and I'm happy to do a zoom. I'll follow up by email.