Expressive Power of Non-deterministic Evolving Recurrent Neural Networks in Terms of Their Attractor Dynamics

Details

Serval ID
serval:BIB_2B6B8310D6DF
Type
Inproceedings: an article in a conference proceedings.
Collection
Publications
Institution
Title
Expressive Power of Non-deterministic Evolving Recurrent Neural Networks in Terms of Their Attractor Dynamics
Title of the conference
Unconventional Computation and Natural Computation: 14th International Conference
Author(s)
Cabessa J., Duparc J.
Publisher
Springer, Cham
Address
Auckland, New Zealand
ISBN
978-3-319-21818-2
978-3-319-21819-9
Publication state
Published
Issued date
2015
Peer-reviewed
Oui
Volume
9252
Series
Lecture Notes in Computer Science
Pages
144-156
Language
english
Notes
UCNC 2015, , August 30 - September 3, 2015, Proceedings
Abstract
We introduce a model of nondeterministic hybrid recurrent neural networks – made up of Boolean input and output cells as well as internal sigmoid neurons, and equipped with the possibility to have their synaptic weights evolve over time, in a nondeterministic manner. When subjected to some infinite input stream and some specific synaptic evolution, the networks necessarily exhibit some attractor dynamics in their Boolean output cells, and accordingly, recognize some specific neural ω -languages. The expressive power of these networks is measured via the topological complexity of their underlying neural ω -languages. In this context, we prove that the two models of rational-weighted and real-weighted nondeterministic hybrid neural networks are computationally equivalent, and recognize precisely the set of all analytic neural ω -languages. They are therefore strictly more expressive than the nondeterministic Büchi and Muller Turing machines.
Keywords
Recurrent neural networks, Neural computation, Analog computation, Evolving systems, Attractors, Turing machines, Expressive power
Create date
10/05/2017 13:22
Last modification date
21/08/2019 5:18
Usage data