Abstract
In this paper some new techniques to analyze nonlinear neural networks are reviewed. A neural network is called nonlinear if the introduction of new data into the synaptic efficacies has to be performed through a non-linear operation. The original Hopfield model is linear whereas, for instance, clipped synapses constitute a nonlinear model. We examine the statistical mechanics of a nonlinear neural network with finitely many patterns and arbitrary synaptic kernel, study the information retrieval, and show how the abundantly present spurious states which are a consequence of the nonlinearity can be eliminated.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Preview
Unable to display preview. Download preview PDF.
References
J. J. Hopfield, Proc. Natl. Acad. Sci. U.S.A. 79 (1982) 2554
See D. J. Amit, these proceedings, for details
W. C. McCulloch and W. Pitts, Bull. Math. Biophys. 5 (1943) 115
P. Peretto, Biol. Cybernet. 50 (1984) 51
R. G. Palmer, Adv. Phys. 31 (1982)669
A. C. D. van Enter and J. L. van Hemmen, Phys. Rev. A 29 (1984) 355
D. O. Hebb, The Organization of Behavior (Wiley, New York, 1949)
J. L. van Hemmen and R. Kühn, Phys. Rev. Lett. 57 (1986) 913
H. Sompolinsky, Phys. Rev. A 34 (1986) 2571
J. L. van Hemmen, D. Grensing, A. Huber and R. Kühn, Nonlinear neural networks: I. General theory, II. Information processing, to be published.
D. Grensing, R. Kühn and J. L. van Hemmen, J. Phys. A.: Math. Gen. 20 (1987)
J. L. van Hemmen, Phys. Rev. Lett. 49 (1982) 409
J. L. van Hemmen, D. Grensing, A. Huber and R. Kühn, Z. Phys. B 65 (1986) 53
J. Lamperti, Probability (Benjamin, New York, 1966) Sec. 7. More generally, one needs ergodicity.
D. Grensing and R. Kühn, J. Phys. A: Math. Gen 19 (1986) L 1153
J. L. van Hemmen, in: Heidelberg Colloquium on Spin Glasses, edited by J. L. van Hemmen and I. Morgenstern, Lecture Notes in Physics 192 (Springer, Berlin, 1983) pp. 203–233, in particular the Appendix.
The intuitive justification of the Laplace argument is clear. For a proof, see: N. G. de Bruyn, Asymptotic Methods in Analysis, 2nd Ed. (North-Holland, Amsterdam, 1961) Sec. 4.2
E. T. Copson, Asymptotic Expansions (Cambridge University Press, Cambridge, 1965) Chap. 5
J. L. van Hemmen, Phys. Rev. A 34 (1986) 3435
This convention differs from the one in Ref. 8. There λ denotes an eigenvalue of QP where P = diag (pγ).
G. Iooss and D. D. Joseph, Elementary Stability and Bifurcation Theory (Springer, Berlin, 1980)
J. J. Hopfield, in: Modelling and Analysis in Biomedicine, edited by C. Nicolini (World Scientific, Singapore, 1984) pp. 369–389, in particular p. 381
G. Parisi, J. Phys. A: Math. Gen. 19 (1986) L 617
J. P. Nadal, G. Toulouse, J.-P. Changeux, and S. Dehaene, Europhys. Lett. 1 (1986) 535
D. J. Amit, H. Gutfreund, and H. Sompolinsky, Phys. Rev. A 32 (1985) 1007
K. L. Chung, Elementary Probability Theory and Stochastic Processes, 2nd Ed. (Springer, Berlin 1975) Eqs. (7.3.5) - (7.3.7) on pp. 211–212.
Since sgn (x) is odd, λρ vanishes for | ρ | even; cf. (3.6). About half of the remaining eigenvalues is positive.
See Ref. 10. An upper bound is provided by q-1.
J.-P. Changeux, S. Dehaene and G. Toulouse, Proc. Natl. Acad. Sci. U.S.A. 83 (1986) 1695
J. L. van Hemmen and K. Rzążewski, to be published
J. L. van Hemmen and A. C. D. van Enter, Phys. Rev. A 34 (1986) 2509
J. L. van Hemmen, A. C. D. van Enter, and J. Canisius, Z. Phys. B 50 (1983) 311
Interestingly, states which are products of two patterns bifurcate first. This allows logical operations such as EQUIVALENCE.
Author information
Authors and Affiliations
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 1987 Springer-Verlag Berlin Heidelberg
About this paper
Cite this paper
van Hemmen, J.L. (1987). Pattern Recognition in Nonlinear Neural Networks. In: Güttinger, W., Dangelmayr, G. (eds) The Physics of Structure Formation. Springer Series in Synergetics, vol 37. Springer, Berlin, Heidelberg. https://doi.org/10.1007/978-3-642-73001-6_2
Download citation
DOI: https://doi.org/10.1007/978-3-642-73001-6_2
Publisher Name: Springer, Berlin, Heidelberg
Print ISBN: 978-3-642-73003-0
Online ISBN: 978-3-642-73001-6
eBook Packages: Springer Book Archive