BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:The predictive power of contractive neural-network spaces - Ashley
  Mills
DTSTART:20050405T153000Z
DTEND:20050405T163000Z
UID:TALK4306@talks.cam.ac.uk
CONTACT:Phil Cowans
DESCRIPTION:Prediction power comes from being able to identify similar sit
 uations \nthrough time and associate with those points the generalised not
 ions of what \nhappens next according to observation. Positive state separ
 ability\, with \nrespect to the transient dynamics of time series\, is tha
 t which boosts the \ndifferences between temporal artifacts which are prec
 ursors and contributors \nto events of subsequent eminent importance that 
 one would like to predict.\n\nThe idea of inducing linear separability of 
 a classification space through \nforced dimensionality explosion is not ne
 w. In the field of recurrent neural \nnetworks\, realisations of a tempora
 l variant of this paradigm have been \ncropping up with increasing frequen
 cy\, with proponents boasting their \nsuperior classification power and en
 gineering simplicity.\n\nSome examples are the echo state network (ESN) fr
 om Herbert Jaeger\, the \nliquid state machine (LSM) from Wolfgang Maass\,
  and the \nback-propagation-de-correlation algorithm (BPDC) and associated
  dynamic \nnetwork from Jochen Steil.\n\nWhilst these models vary in detai
 ls\, abstractly they are very similar. The \nidea is simple\; project a te
 mporal stream into a high-dimensional space to \nincrease its state-separa
 bility\, and using a linear or nearly-linear readout \nmechanism\, exploit
  the enhanced separability to perform prediction and/or \nclassification.\
 n\nAlthough the idea is simple\, the models tend toward an undesirable deg
 ree of \n"kernel magic" to acheive their satisfying performance\; there ha
 s yet to be \na precise and sufficient qualification of the mechanisms inv
 olved.\n\nThis talk will expand on the above notions and discuss the one-s
 emester \nmini-project work experiments which were executed on the hunch t
 hat the \npower of these networks derive principally from fractal contract
 ivity of the \nrecurrent network weight spaces.
LOCATION:Ryle Seminar Room\, Cavendish Laboratory
END:VEVENT
END:VCALENDAR