BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:A normative account of episodic memory in online learning over ope
 n model spaces - Gergo Orban\, MTA Wigner Research Centre for Physics\, Bu
 dapest
DTSTART:20150824T150000Z
DTEND:20150824T160000Z
UID:TALK60504@talks.cam.ac.uk
CONTACT:Guillaume Hennequin
DESCRIPTION: Both the human brain and artificial learning agents operating
  in real-world or comparably complex environments are faced with the probl
 em of online model selection. The reason for this is that both the amount 
 and dimensionality of the data and the dimensionality of the model space i
 s huge or even infinite. In principle this can be handled: hierarchical Ba
 yesian inference gives a principled method for model selection and it conv
 erges on the same posterior for both batch and online learning. However\, 
 maintaining a parameter posterior for each model in parallel has in genera
 l an even higher memory cost than storing the entire data set and is conse
 quently clearly unfeasible. On the other hand\, the sufficient statistic f
 or one model will usually not be sufficient for the fitting of different k
 ind of model meaning that the agent loses information with each model chan
 ge.  We propose that episodic memory can circumvent the challenge of limit
 ed memory-capacity online  model selection by retaining a selected subset 
 of data points.  We design a method to compute the quantities necessary fo
 r model selection even when the data is discarded and only statistics of o
 ne (or few) learnt models are available.  We demonstrate on a simple model
  that a limited-sized episodic memory buffer\, when the content is optimis
 ed to retain data with statistics not matching the current representation\
 , can resolve the fundamental challenges of online model selection.\n
LOCATION:Cambridge University Engineering Department\, CBL\, BE-438 (http:
 //learning.eng.cam.ac.uk/Public/Directions)
END:VEVENT
END:VCALENDAR