Advanced search
1 file | 666.62 KB Add to list

Technical report on hierarchical reservoir computing architectures

(2012)
Author
Organization
Abstract
One approach for building architectures (of which an overview was given in D.6.1) in AMARSi is to use reservoir computing. Here, untrained (or unsupervised trained) recurrent neural networks are used for motion control by learning simple readouts on the dynamic representation generated by the dynamic RNN system. Although single reservoirs are able to generate rich and tunable control patterns (as demonstrated in D.4.1), to allow composition of motion or high-level control, these modules need to be built in an architecture. An active research area in reservoir computing is to build hierarchical reservoir systems. The main reason for this is that reservoirs basically are band-pass systems and can only represent information in a limited frequency band. If information at both fast and slow timescales needs to be integrated, a natural approach is to build a hierarchical system where each layer operates at a different time scale. The big challenge in these hierarchies is how to learn intermediate representations that link the various layers, and especially how bottom-up and top-down information flows need to be organized. We believe that these hierarchical reservoir computing systems are good candidates to build (at least part of) architectures required in AMARSi for rich motor control. In this short deliverable we give an overview of and references to current approaches in hierarchical reservoir computing, several of which have been investigated on speech and handwriting recognition problems in the sister EU project ORGANIC (http://reservoir- computing.org/organic). Many of these hierarchical systems can be used to not only generate dynamical feature hierarchies, but are also able to learn a hierarchy of pattern controller, of special interest to the AMARSi project.

Downloads

  • deliverable5.2.pdf
    • full text
    • |
    • open access
    • |
    • PDF
    • |
    • 666.62 KB

Citation

Please use this url to cite or link to this publication:

MLA
wyffels, Francis, et al. Technical Report on Hierarchical Reservoir Computing Architectures. 2012.
APA
wyffels, F., Waegeman, T., Schrauwen, B., & Jaeger, H. (2012). Technical report on hierarchical reservoir computing architectures.
Chicago author-date
wyffels, Francis, Tim Waegeman, Benjamin Schrauwen, and Herbert Jaeger. 2012. “Technical Report on Hierarchical Reservoir Computing Architectures.”
Chicago author-date (all authors)
wyffels, Francis, Tim Waegeman, Benjamin Schrauwen, and Herbert Jaeger. 2012. “Technical Report on Hierarchical Reservoir Computing Architectures.”
Vancouver
1.
wyffels F, Waegeman T, Schrauwen B, Jaeger H. Technical report on hierarchical reservoir computing architectures. 2012.
IEEE
[1]
F. wyffels, T. Waegeman, B. Schrauwen, and H. Jaeger, “Technical report on hierarchical reservoir computing architectures.” 2012.
@misc{3005080,
  abstract     = {{One approach for building architectures (of which an overview was given in D.6.1) in AMARSi is to use reservoir computing. Here, untrained (or unsupervised trained) recurrent neural networks are used for motion control by learning simple readouts on the dynamic representation generated by the dynamic RNN system. Although single reservoirs are able to generate rich and tunable control patterns (as demonstrated in D.4.1), to allow composition of motion or high-level control, these modules need to be built in an architecture. An active research area in reservoir computing is to build hierarchical reservoir systems. The main reason for this is that reservoirs basically are band-pass systems and can only represent information in a limited frequency band. If information at both fast and slow timescales needs to be integrated, a natural approach is to build a hierarchical system where each layer operates at a different time scale. The big challenge in these hierarchies is how to learn intermediate representations that link the various layers, and especially how bottom-up and top-down information flows need to be organized. We believe that these hierarchical reservoir computing systems are good candidates to build (at least part of) architectures required in AMARSi for rich motor control. In this short deliverable we give an overview of and references to current approaches in hierarchical reservoir computing, several of which have been investigated on speech and handwriting recognition problems in the sister EU project ORGANIC (http://reservoir- computing.org/organic). Many of these hierarchical systems can be used to not only generate dynamical feature hierarchies, but are also able to learn a hierarchy of pattern controller, of special interest to the AMARSi project.}},
  author       = {{wyffels, Francis and Waegeman, Tim and Schrauwen, Benjamin and Jaeger, Herbert}},
  language     = {{eng}},
  title        = {{Technical report on hierarchical reservoir computing architectures}},
  url          = {{https://www.amarsi-project.eu/system/files/deliverable5.2.pdf}},
  year         = {{2012}},
}