search

NeuroDataDesign / fngs-archive

Apache License 2.0
2 stars 1 forks source link

pGMM #112

Open 02agarwalt opened 7 years ago

02agarwalt commented 7 years ago

write pGMM algorithm, do methods_paper.md on it, and write proof of GMM. pGMM = prior-based GMM. w prior GMM, we know what the w matrix is ahead of time, and the pi-vector. Future iterations will constrain the expectation w vector to not vary from the prior we pass in.

jovo commented 7 years ago

@ebridge2 i'm not quite sure what this is. i'd rather you do a "benchmarks", making a single page in "https://github.com/neurodata/checklists/blob/master/methods_paper.md" style, but including the methods that you have already (especially the synthetic data), so we have some things to compare against. what do you think?

ebridge2 commented 7 years ago

I agree; I just wanted to start somewhere more foundational for the tGMM before getting back there. As of now, the tGMM still does not converge in many cases, and I think it is a flaw in the algorithm itself, so what I was going to do was go to a much simpler model first to modify (we went from regular GMM -> tGMM; I think doing a pGMM before finishing tGMM will help us understand how to "better" do the tGMM in the next iteration, because I think that algorithm is one step ahead of us from a conceptual understanding right now). My goal for Magnetrons already is to finish that synthetic trial to be replicated on all 3; maybe a tbd by 2 weeks could be to have all 3 (wwCon, Null model, and pGMM) combined into one cohesive report, and then have a model due for tGMM++ as well, and then write up a methods_algorithms.md the following week (ie, 3 weeks out)?