Home

About

Advanced Search

Browse by Discipline

Scientific Societies

E-print Alerts

Add E-prints

E-print Network
FAQHELPSITE MAPCONTACT US


  Advanced Search  

 
LITDOCS/630077.1 The Neocortex as a Hebbian Proofreader
 

Summary: LITDOCS/630077.1
The Neocortex as a Hebbian Proofreader
Kingsley J. A. Cox, Anca Radulescu and Paul R. Adams
Department of Neurobiology and Behavior, College of Arts and Sciences, SUNY Stony Brook, State University of
New York, Stony Brook, NY 11794-5230, USA
Abstract
We propose that the neocortex is a machine for learning high order correlations which avoids error
catastrophes induced by relatively unstructured environments using an online Hebbian proofreading
canonical microcircuit. We argue that there is inevitable crosstalk between individual weight updates, for
example caused by spine-spine calcium spillover. This reflects the incompatible requirements for voltage
spread and calcium localization in networks that use voltage to compute and calcium to learn; the ratio of
their space constants sets a plasticity error level. We show that for linear neurons which only learn
pairwise statistics such errors do no completely prevent selforganisation, although learning gets worse as
the network enlarges. But if neurons are nonlinear (essential for learning high-order statistics), a learning
error collapse occurs at a network size given by the reciprocal error rate, around 1 thousand. Since the
cortex cannot know in advance what environments could trigger a collapse, and the column size is around
1000 neurons, it must use an independent online assessment of current input statistics to gate learning by
feedforward networks, a form of Hebbian proofreading. We show that a microcircuit which guarantees
catastrophe avoidance and which would therefore the allow neocortex to act as a universal learning
machine closely resembles the physiology and anatomy of layer 6 cells and connections. Proofreading

  

Source: Adams, Paul R. - Department of Neurobiology and Behavior, SUNY at Stony Brook

 

Collections: Biology and Medicine