Abstract
In this paper, the derivative of the input scaling and spectral radius parameters of Echo State Network reservoir are derived. This was achieved by re-writing the reservoir state update equation in terms of template matrices whose eigenvalues can be pre-calculated, so the two parameters appear in the state update equation in the form of simple multiplication which is differentiable. After that the paper derives the derivatives and then discusses why direct application of these two derivatives in gradient descent to optimize reservoirs in a sequential manner would be ineffective due to the nature of the error surface and the problem of large eigenvalue spread on the reservoir state matrix. Finally it is suggested how to apply the derivatives obtained here for joint-optimizing the reservoir and readout at the same time.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Notes
- 1.
To avoid confusion, we shall refer to the process of choosing good reservoir parameter as “tuning” of ESN, while “training” means to adapt the weights of the readout layer, given some reservoir.
- 2.
The spectral radius of a matrix is the maximum of the magnitudes of its eigenvalues.
- 3.
If the sparseness is above 90 % eigenvalue calculation for \( \mathbf {W} \) may sometime fail due to numerical problems (using Python’s Scipy package).
- 4.
The desired response is needed for supervised sequential learning. It is defined by what filter configuration the ESN is to be operated in. For details, see [1].
- 5.
It is disjoint because, for each point (\( s, \tilde{\rho } \)), \( \mathbf {w}_{\text {out}} \) is solved for by the method of least squares. Since least squares involves matrix inverse, for a slightly different point \( s+\varDelta s,\tilde{\rho } + \varDelta \tilde{\rho } \), a very different \( \mathbf {w}_{\text {out}} \) may be produced.
References
Haykin, S.S.: Adaptive Filter Theory, 4th edn. Pearson Education India, New Delhi (2005)
Jaeger, H.: The “echo state” approach to analysing and training recurrent neural networks. Technical report GMD Report 148, German National Research Center for Information Technology (2001)
Jaeger, H.: Tutorial on training recurrent neural networks, covering BPPT, RTRL, EKF and the “echo state network” approach. GMD-Forschungszentrum Informationstechnik (2002)
Jaeger, H.: Adaptive nonlinear system identification with echo state networks. Networks 8, 9 (2003)
Jaeger, H.: Reservoir riddle: suggestions for echo state network research. In: Proceedings of International Joint Conference on Neural Networks, pp. 1460–1462 (2005)
Jaeger, H., Haas, H.: Harnessing nonlinearity: predicting chaotic systems and saving energy in wireless communication. Science 304(5667), 78–80 (2004)
Jiang, F., Berry, H., Schoenauer, M.: Supervised and evolutionary learning of echo state networks. In: Rudolph, G., Jansen, T., Lucas, S., Poloni, C., Beume, N. (eds.) PPSN 2008. LNCS, vol. 5199, pp. 215–224. Springer, Heidelberg (2008)
Küçükemre, A.U.: Echo state networks for adaptive filtering. Ph.D. thesis, University of Applied Sciences (2006)
Lukoševičius, M.: A practical guide to applying echo state networks. In: Montavon, G., Orr, G.B., Müller, K.-R. (eds.) Neural Networks: Tricks of the Trade, 2nd edn. LNCS, vol. 7700, pp. 659–686. Springer, Heidelberg (2012)
Lukoševičius, M., Jaeger, H.: Reservoir computing approaches to recurrent neural network training. Comput. Sci. Rev. 3(3), 127–149 (2009)
Petersen, K.B., Pedersen, M.S., et al.: The matrix cookbook. Technical University of Denmark, vol. 7, p. 15 (2008)
Schrauwen, B., Verstraeten, D., Van Campenhout, J.: An overview of reservoir computing: theory, applications and implementations. In: Proceedings of the 15th European Symposium on Artificial Neural Networks, pp. 471–482 (2007)
Williams, R.J., Zipser, D.: A learning algorithm for continually running fully recurrent neural networks. Neural Comput. 1(2), 270–280 (1989)
Xia, Y., Jelfs, B., Van Hulle, M.M., Príncipe, J.C., Mandic, D.P.: An augmented echo state network for nonlinear adaptive filtering of complex noncircular signals. IEEE Trans. Neural Netw. 22(1), 74–83 (2011)
Yuenyong, S.: Fast and effective tuning of echo state network reservoir parameters using evolutionary algorithms and template matrices. In: 19th International Computer Science and Engineering Conference (ICSEC), November 2015
Yuenyong, S., Nishihara, A.: Evolutionary pre-training for CRJ-type reservoir of echo state networks. Neurocomputing 149, 1324–1329 (2015)
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2016 Springer International Publishing Switzerland
About this paper
Cite this paper
Yuenyong, S. (2016). On the Gradient-Based Sequential Tuning of the Echo State Network Reservoir Parameters. In: Booth, R., Zhang, ML. (eds) PRICAI 2016: Trends in Artificial Intelligence. PRICAI 2016. Lecture Notes in Computer Science(), vol 9810. Springer, Cham. https://doi.org/10.1007/978-3-319-42911-3_54
Download citation
DOI: https://doi.org/10.1007/978-3-319-42911-3_54
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-319-42910-6
Online ISBN: 978-3-319-42911-3
eBook Packages: Computer ScienceComputer Science (R0)