BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Neuronal processing of continuous sensory streams - Robert Gütig\
 , Max Planck Institute for Experimental Medicine\, Göttingen\, Germany
DTSTART:20131108T111500Z
DTEND:20131108T121500Z
UID:TALK48747@talks.cam.ac.uk
CONTACT:Guillaume Hennequin
DESCRIPTION:During behavior\, a continuous stream of sensory information r
 eaches the central nervous system in the form of a high-dimensional spatio
  temporal pattern of action potentials. When processing such activity\, ma
 ny sensory neurons respond with high selectivity and precise tuning to beh
 aviorally meaningful sensory cues\, such as sounds within a communication 
 call or shapes within an evolving visual scene. Typically\, the temporal e
 xtent of such embedded features can be orders of magnitude shorter than th
 e duration of the encompassing behavioral episodes. It is unclear how neur
 ons can bridge between these time scales when learning to signal to downst
 ream processing stages the presence and quality of individual perceptual f
 eatures in an evolving sensory stream. It is commonly hypothesized that su
 ch learning must rely on temporal segmentation of the sensory streams or o
 ther types of supervisory signals that provide neurons with information ab
 out the timing and values of their target features. In contrast\, we show 
 here that an aggregate scalar teaching signal delivered at the and of a lo
 ng sensory episode is sufficient for biologically plausible neuron models 
 to acquire even complex tuning functions for spatio-temporal patterns of\n
 spikes that arrive embedded in continuous streams of spiking background ac
 tivity. The proposed learning implements a novel form of spike-based synap
 tic plasticity that reduces the difference between a neuron's output spike
  count and the value of the teaching signal. Based on the simplicity of su
 ch supervisory signaling we propose a novel type of self-organizing spikin
 g neuronal networks as a model for the emergence of feature selectivity an
 d map formation in sensory pathways. In these two-layer networks\, self-or
 ganization is driven by a positive feedback loop between a processing and 
 a supervisor layer that computes neuronal teaching signals by taking weigh
 ted averages over the processing layer activity. We demonstrate the power 
 of this learning paradigm by implementing a neuronal model of continuous s
 peech processing.
LOCATION:Cambridge University Engineering Department\, CBL Rm #438 (http:/
 /learning.eng.cam.ac.uk/Public/Directions)
END:VEVENT
END:VCALENDAR