BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Computational Neuroscience Journal Club - Michael Rule (Control Gr
 oup)
DTSTART:20181113T160000Z
DTEND:20181113T170000Z
UID:TALK114829@talks.cam.ac.uk
CONTACT:Rodrigo Echeveste
DESCRIPTION:Michael Rule will be presenting:\n\n• Alleviating catastroph
 ic forgetting using context-dependent gating and synaptic stabilization\n\
 n• Nicolas Y. Masse\, Gregory D. Grant\, and David J. Freedman\n\n• PN
 AS (2018)\n\n• http://www.pnas.org/content/115/44/E10467\n\n\nAbstract: 
 Humans and most animals can learn new tasks without forgetting old ones. H
 owever\, training artificial neural networks (ANNs) on new tasks typically
  causes them to forget previously learned tasks. This phenomenon is the re
 sult of “catastrophic forgetting\,” in which training an ANN disrupts 
 connection weights that were important for solving previous tasks\, degrad
 ing task performance. Several recent studies have proposed methods to stab
 ilize connection weights of ANNs that are deemed most important for solvin
 g a task\, which helps alleviate catastrophic forgetting. Here\, drawing i
 nspiration from algorithms that are believed to be implemented in vivo\, w
 e propose a complementary method: adding a context-dependent gating signal
 \, such that only sparse\, mostly nonoverlapping patterns of units are act
 ive for any one task. This method is easy to implement\, requires little c
 omputational overhead\, and allows ANNs to maintain high performance acros
 s large numbers of sequentially presented tasks\, particularly when combin
 ed with weight stabilization. We show that this method works for both feed
 forward and recurrent network architectures\, trained using either supervi
 sed or reinforcement-based learning. This suggests that using multiple\, c
 omplementary methods\, akin to what is believed to occur in the brain\, ca
 n be a highly effective strategy to support continual learning.
LOCATION:Cambridge University Engineering Dept.\, Board Room (2nd floor\, 
 Baker building)
END:VEVENT
END:VCALENDAR