Incremental Associative Learning

Erik Jonsson, Michael Felsberg, Gösta Granlund
Research Report
LiTH-ISY-R-2691
Sept, 2005

Abstract

The channel representation is a biologically inspired information representation used in e.g. learning (associative networks) and image denoising (channel smoothing). An efficient incremental version of the previous batch mode learning algorithms is presented, that can also update separate training and validation sets in parallel. The performance of the different methods is studied on a simple example. Furthermore, the associative networks are used to learn the Hough transform voting function, and it is shown under which circumstances the network output can be interpreted as a conditional probability density. The behavior of the learned voting function using the different methods is examined experimentally.

Bibtex entry

@techreport{jfg05,
        Author = {Erik Jonsson and Michael Felsberg and G{\"o}sta Granlund},
        Date-Added = {2006-08-08 13:43:36 +0200},
        Date-Modified = {2007-08-07 11:15:11 +0200},
        Institution = {Dept. EE, Link\"oping University},
        Month = {Sept},
        Number = {LiTH-ISY-R-2691},
        Title = {Incremental Associative Learning},
        Year = {2005}}