Publication View

Further Results on Controllability of Recurrent Neural Networks (1999)

Abstract
This paper studies controllability properties of recurrent neural networks. The new contributions are: (1) an extension of a previous result to a slightly different model, (2) a formulation and proof of a necessary and sufficient condition, and (3) an analysis of a low-dimensional case for which the hypotheses made in previous work do not apply. 1 Introduction This paper deals with controllability properties of what are often called "recurrent neural networks". These constitute a class of nonlinear systems which, although formally analogous to linear systems, exhibit interesting nonlinear characteristics and arise often in applications, see e.g. [3, 4, 5, 7, 9, 10, 11, 14, 15, 18]. A general model of recurrent nets (see e.g. [15]) is as follows. Assume given a Lipschitz map ` : R ! R. The most typical choice is `(x) = tanh x : R ! R : x 7! e x \Gamma e \Gammax e x + e \Gammax ; which is also called the "sigmoid" or "logistic" map. For each positive integer n, we let ~ ` ...

Publication details
Download http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.50.8809
Source http://www.math.rutgers.edu/~sontag/FTP_DIR/qiao-scl.ps.gz
Contributors CiteSeerX
Repository CiteSeerX - Scientific Literature Digital Library and Search Engine (United States)
Type text
Language English
Relation 10.1.1.48.3536, 10.1.1.50.8970, 10.1.1.53.4648, 10.1.1.49.6388, 10.1.1.49.2977, 10.1.1.56.8773