BEGIN:VCALENDAR
VERSION:2.0
PRODID:talks.ox.ac.uk
BEGIN:VEVENT
SUMMARY:A mathematical theory of semantic development in deep neural netwo
rks - Dr Andrew Saxe (Harvard)
DTSTART;VALUE=DATE-TIME:20191022T150000
DTEND;VALUE=DATE-TIME:20191022T160000
UID:https://talks.ox.ac.uk/talks/id/415e7089-f358-4ddc-bbb0-f4e4679fa02a/
DESCRIPTION:An extensive body of empirical research has revealed remarkabl
e regularities in the acquisition\, organization\, deployment\, and neural
representation of human semantic knowledge\, thereby raising a fundamenta
l conceptual question: what are the theoretical principles governing the a
bility of neural networks to acquire\, organize\, and deploy abstract know
ledge by integrating across many individual experiences? I will describe
work addressing this question by mathematically analyzing the nonlinear dy
namics of learning in deep linear networks. We find exact solutions to th
is learning dynamics that yield a conceptual explanation for the prevalenc
e of many disparate phenomena in semantic cognition\, including the hierar
chical differentiation of concepts through rapid developmental transitions
\, the ubiquity of semantic illusions between such transitions\, the emerg
ence of item typicality and category coherence as factors controlling the
speed of semantic processing\, changing patterns of inductive projection o
ver development\, and the conservation of semantic similarity in neural re
presentations across species. Furthermore\, I will show that many of these
phenomena only arise in deep but not shallow networks. Thus\, surprisingl
y\, this simple neural model qualitatively recapitulates many diverse regu
larities underlying semantic development\, while providing analytic insigh
t into how the statistical structure of an environment can interact with n
onlinear deep learning dynamics to give rise to these regularities.\nSpeak
ers:\nDr Andrew Saxe (Harvard)
LOCATION:New Radcliffe House (Room 20.05 Seminar Room 2)\, Walton Street O
X2 6NW
URL:https://talks.ox.ac.uk/talks/id/415e7089-f358-4ddc-bbb0-f4e4679fa02a/
BEGIN:VALARM
ACTION:display
DESCRIPTION:Talk:A mathematical theory of semantic development in deep neu
ral networks - Dr Andrew Saxe (Harvard)
TRIGGER:-PT1H
END:VALARM
END:VEVENT
END:VCALENDAR