Summary
International Symposium on Nonlinear Theory and its Applications
2008
Session Number:A2L-C
Session:
Number:A2L-C2
Pattern classification with CNNs as reservoirs
D. Verstraeten, S. Xavier-de-Souza, B. Schrauwen, J. Suykens, D. Stroobandt, J. Vandewalle,
pp.-
Publication Date:2008/9/7
Online ISSN:2188-5079
Summary:
Reservoir Computing is a novel method in the field of neural networks and machine learning, which combines the computational power of a nonlinear dynamic system with the ease of training of a linear classifier. The basic setup is as follows: a sufficiently complex network of nonlinear nodes (called the reservoir) is excited by an input signal, and the instantaneous dynamic response of the system is then used to train a simple linear readout function. In this contribution, we present a proof-of-concept system that demonstrates the possibility of using the nonlinear spatiotemporal dynamics of a Cellular Neural/Nonlinear Network (CNN) to play the role of a reservoir. We discuss the advantages and limitations of this approach and illustrate the idea by using the system to solve both a simple academic task and a real world speech recognition problem. We use a global optimization method called Coupled Simulated Annealing (CSA) to optimize CNN templates that give suitable reservoir properties to the CNN. Finally, we validate the simulation results using an ACE16k CNN chip.