Information processing with recurrent dynamical systems: theory and experiment
Dates : Wednesday 5 September 2012
Importante Note One:
To attend the Satellite Meeting, it is mandatory to register to the European Conference on Complex Systems 2012 ECCS2012.
Although computer science and physics are two distinct fields, they are intimately linked. “Information is physical” (R. Landauer) and must be processed by physical devices. This fact is behind such far reaching conclusions as the Church-Turing thesis (the statement that all classical computers have equivalent computational power), the resolution of Maxwell’s paradox, the introduction of quantum computing, or the goal of understanding how information is processed in biological neural networks.
Over the past 10 years the machine learning community has introduced the concept of “reservoir computing”, which provides another such connection. At the heart of a reservoir computer is the “reservoir”, a dynamical system of the type studied by physicists interested in complex systems, composed of many internal variables that exhibit a transition from stability to deterministic chaos. To transform such a dynamical system into a reservoir computer one adds to it an input layer in order to drive the dynamical system with the time dependent signal to be analyzed, and an output layer which carries out simple and “learnable” post-processing to produce an output which is based on the internal states of the reservoir. Remarkably, this simple scheme provides state of the art performance for tasks which are generally deemed hard, such as speech recognition or time series prediction.
This new paradigm provides a connection between physics, computer science and engineering. From the physics point of view, reservoir computing raises new and fascinating questions: What is the computational power of nonlinear dynamical systems? How can one characterize the computational power? Can the powerful tools of theoretical physics and information theory be brought to bear on this question? From the point of view of engineering, another fascinating question arises: Can the flexibility of the reservoir computing paradigm be used to devise new architectures for computing machines? From the point of view of computer science, one can hope that a composable or even programmable approach allows solving more complex tasks by combining multiple reservoir computing modules.
Some exciting results at the interface of these three fields have just appeared, including the experimental realization of reservoirs with performance, not only in error on a given task, but also in speed, comparable to the best digital implementations. We believe that these are the first steps towards what will become a very dynamic, interdisciplinary, research area.
The aim of this satellite meeting is to bring together researchers working at the interface between these three areas, in order to foster the development of this new research area. Questions of interest include:
-novel/improved methods to train recurrent dynamical systems for information processing
-theoretical tools for understanding the information processing capability of recurrent dynamical systems
-experimental implementations of reservoir computers, including links to unconventional architectures (for instance relation to robotics and morphological computing)
Serge Massar(Laboratoire d’Information Quantique, ULB)
Benjamin Schrauwen (Reservoir Lab,Universiteit Gent)
Ingo Fischer (Institute for Cross-Disciplinary Physics and Complex Systems, Universitat de les Illes Balears)
For the Program of the Satellite Meeting click HERE.
For presentations given at the meeting, click HERE
For photographs from the meeting, click HERE