To be discussed with AD, YW

tlienart / thox

[done] phd thesis @ oxford stats

0 stars 0 forks source link

Closed tlienart closed 7 years ago

tlienart commented 7 years ago

TFS section --> notion of consistency (see also email sent to AD)
~EPBP lack of assumptions --> we still need to sample from terms of form \psi_st don't we? see #12~ Actually no since we only sample through q_s but careful thought

tlienart commented 7 years ago

~link between EPBP and TFS is that in EPBP we assume we can sample from \psi_st(X^(i)_s, x_t) for all pairs of node. In HMM, this would require to also be able to sample from the backward dynamic p(x_t | x_{t+1}) which is not the case in NLG model. This should definitely be touched upon and the link clarified.~ No sampling of psi_st required. Only of q (which is in expoF) this should also mean that EPBP can in fact be used for smoothing.