BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Model Selection and Local Geometry
DTSTART:20220520T151500
DTEND:20220520T170000
DTSTAMP:20260406T184228Z
UID:aed27568d268b685e453f70ff90475a683cff8e05913c146d41b3c6f
CATEGORIES:Conferences - Seminars
DESCRIPTION:Prof. Robin Evans\, Oxford University\nWe consider problems in
  model selection caused by the geometry of models close to their points o
 f intersection.  In some cases---including common classes of causal or 
 graphical models\, as well as time series models—distinct models may n
 evertheless have identical tangent spaces.  This has two immediate cons
 equences: first\, in order to obtain constant power to reject one model i
 n favour of another we need local alternative hypotheses that decrease to
  the null at a slower rate than the usual parametric $n^{-1/2}$ (typicall
 y we will require $n^{-1/4}$ or slower)\; in other words\, to distinguish
  between the models we need large effect sizes or very large sample sizes
 .  Second\, we show that under even weaker conditions on their tangent 
 cones\, models in these classes cannot be made simultaneously convex by a
  reparameterization.\n\nThis shows that Bayesian network models\, amongst 
 others\, cannot be learned directly with a convex method similar to the g
 raphical lasso. However\, we are able to use our results to suggest metho
 ds for model selection that learn the tangent space directly\, rather tha
 n the model itself.  In particular\, we give a generic algorithm for le
 arning Bayesian network models.\n\nReference\nEvans\, R.J. Model selectio
 n and local geometry\, Annals of Statistics\, 48 (6)\, pp 3514-3544.\n\n
  
LOCATION:MA A1 10 https://plan.epfl.ch/?room==MA%20A1%2010 https://epfl.zo
 om.us/j/66136073806
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR