BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Towards a theory of layered neural circuit architectures - Alireza
  Alemi (École Normale Supérieure)
DTSTART:20160531T103000Z
DTEND:20160531T113000Z
UID:TALK66426@talks.cam.ac.uk
CONTACT:Daniel McNamee
DESCRIPTION:A main challenge in neuroscience is finding a general computat
 ional principle that explains why cortical circuits are organized in parti
 cular structures. I will start out with the optimal storage principle as a
  guideline to derive optimal neural architecture. For optimal storage\, on
 e needs to have the maximal capacity of a neural network and a learning ru
 le to achieve the capacity. For conventional recurrent neural networks\, t
 he maximal capacity is known as the Gardner bound\, and this bound is achi
 eved via the Three-Threshold Learning Rule (3TLR). However\, calculating t
 he storage capacity of hierarchical neural circuits has been problematic. 
 I will present my recent results suggesting that the capacity of an expans
 ive autoencoder increases superlinearly with the expansion ratio using sim
 ulations and Gardner’s replica theory. I will discuss some of the theore
 tical challenges and limitations of these networks.
LOCATION:Cambridge University Engineering Department\, CBL\, BE-438 (http:
 //learning.eng.cam.ac.uk/Public/Directions)
END:VEVENT
END:VCALENDAR