Recursive artificial neural sequence models

Lecture
Datum: 
25. November 2011
Beginn: 
10:00
Raum: 
Q1-136

Abstract

Sequence data are naturally resulting from process observations, chromosome studies, natural language processing, motor skill recordings, and general streaming applications. Representation of sequential context is an important prerequisite for solving problems like classification or predictive regression. A particular challenge is faced when the sequence elements are multivariate vectors. Self-organizing neural networks allow to train vector data representations in a faithful and biologically meaningful way, but special architectures are needed for recursive element processing, i.e. for a context-aware processing of sequential data. The talk will give an introduction to the field and present a neural model in which data are merged with a recursive definition of their context. Properties and applications of such merge self-organizing maps will be provided.