BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY: Biological learning and memory from a theoretical perspective - M
 ate Lengyel (Engineering Department\, University of Cambridge)
DTSTART:20080521T133500Z
DTEND:20080521T142000Z
UID:TALK11842@talks.cam.ac.uk
CONTACT:Danielle Stretch
DESCRIPTION: The sole purpose of having memories about the past is to aid 
 adaptive\nbehaviour in the future. I will present recent work formalizing 
 this\nnormative idea mathematically and employing it to understand two\ndi
 fferent memory systems in the brain.\n\nLearning about the joint statistic
 s of multiple environmental\nvariables is the basis for making valid predi
 ctions. Traditionally\,\nlearning has been described as being achieved thr
 ough storing pairwise\nassociations between variables. However\, from a th
 eoretical\nperspective\, this seems to be clearly sub-optimal: it leads to
 \nrepresenting only second-order correlations of the available\ninformatio
 n while the statistics of our environments are much more\nrichly structure
 d. Bayesian statistics offers a principled solution to\nselecting the opti
 mal complexity of a representation. Thus\, we\ninvestigated whether human 
 learning is best described by simple\nassociative or more sophisticated Ba
 yesian learning in a visual chunk\nlearning paradigm. We found that human 
 subjects can learn about the\nstatistics of visual stimuli in a highly eff
 icient way\, close to the\noptimum defined by Bayesian model comparison\, 
 which surpasses what was\npredicted by conventional pairwise associative t
 heories.\n\nIn light of the efficiency of this form of 'semantic' learning
 \, it is\npuzzling why we have other memory systems\, such as the one that
  stores\nindividual autobiographical episodes? To answer this question\, r
 ather\nthan considering single isolated predictions\, we have studied the 
 use\nof memories for sequential decision making\, when feedback (in the fo
 rm\nof rewards) is delayed. We show that under specific\, and behaviourall
 y\nrelevant\, conditions a system storing episodes can best a system using
 \n'semantic' memories\, storing the sufficient statistics of the\nenvironm
 ent\, because its relative inefficiency to represent\ninformation about th
 e environment is compensated for by the accuracy\nof decisions that can be
  based on it.  \n\nThe brain has a unique capacity to learn continuously a
 bout the\nenvironment and to use this knowledge flexibly to make predictio
 ns and\nguide its future decisions. Mathematical theories can provide a\np
 owerful tool for dissecting the impressive complexities underlying\nthis r
 emarkable feat.
LOCATION:MR2\, Centre for Mathematical Sciences\, Wilberforce Road\, Cambr
 idge
END:VEVENT
END:VCALENDAR