Open fritzo opened 5 years ago
cc @martinjankowiak
ah, thanks, i was unaware of some of those. added [6]
@fehiepsi Do you have time to look into a pyro.ops.studentt analogous to pyro.ops.gaussian? I think a pyro.distributions.StudentTHMM
would be a very competitive time series model if we can make the math work (say with common dof for all transition and emission distributions).
Note I'm not sure how much is possible, e.g. maybe we can only support univariate Student t.
@fritzo I have some discussions with @martinjankowiak previously about MVStudentT. Hope that it already settles all the math. I will see if it works in a few days (currently I am working on square root version of Gaussian).
aah unfortunately all those conversations were in the context of variational interference (and so not directly applicable)
On Mon, Aug 19, 2019, 6:04 PM Du Phan notifications@github.com wrote:
@fritzo https://github.com/fritzo I have some discussions with @martinjankowiak https://github.com/martinjankowiak previously about MVStudentT. Hope that it already settles all the math. I will see if it works in a few days (currently I am working on square root version of Gaussian).
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/pyro-ppl/funsor/issues/190?email_source=notifications&email_token=AFIUZAXN627SPTOQDFZRGRLQFLABTA5CNFSM4IKN22N2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD4TOSNY#issuecomment-522643767, or mute the thread https://github.com/notifications/unsubscribe-auth/AFIUZAS34YPD6ZEXGETW3STQFLABTANCNFSM4IKN22NQ .
This looks non-trivial but seems doable and interesting. I'll take care of implementing studentt
but it might take a bit of time for me to settle all the math (especially to see if we can derive an information form for the filter in reference [3]).
I think we could do this with an auxiliary variable reparameterizer like pyro.poutine.reparam
. In funsor I believe this would be a reparam
interpretation that could e.g. transform a StudentT
or Stable
to a Joint
with other variables. I doubt we yet have the machinery for an end-to-end heavy-tailed HMM example, but that could probably be done in ~1 week of coding. See Pyro's Forecasting II example.
While Gaussian models are flexible and tractable, real data often exhibits noise with heavier tails than Gaussian. A popular model for robust estimation is Student's t distribution, e.g. as used in Slawek Smyl's models. Like the Gaussian, Student's t filters [1,2,3] and processes [4,5] with shared dof parameter are composable in exact inference. Moreover approximate algorithms allow for differing dof parameters in different parts of the model [3].
References