Abstract
Starting from prescriptions for the learning and retrieval of correlated time-independent patterns in spin glass models we present a class of learning rules for the implementation of temporal sequences of patterns into spin models with randomly delayed interactions. From biological motivation a learning process is devised in such a way that, while learning, the spin-dynamics and the dynamics in the interactions are coupled and take place simultaneously. After training, such a system allows good retrieval of strongly noisy versions of complex sequences of highly correlated patterns such as words considered to be temporal sequences of letters, without making use of multispin coupling. However, the storage of given sequences as stable limit cycles of the network remains in many cases somewhat imperfect during the considered learning time. The small number of errors which are left, lead to a gradually decreasing recognition quality unless the noisy sequence to be recognized is continuously presented to the system with small amplitude via additional receptors. By the use of local receptor fields one may also recognize unlearnt combinations of fragments of the learnt sequences. Perfect learning and retrieval of such sequences is guaranteed under certain conditions by a coupling matrix analogous to the pseudoinverse solution [11] for static patterns. Two different types of networks are investigated in computer simulations of finite systems: (I) a one-layered fully connected network and (II) an extension of a so-called index model proposed by Schmitz et al. [26] for time-independent patterns, which consists of two layers of neurons with connections only inbetween the layers. In the second model the transitions between two patterns of a sequence are perfectly sharp, one reason why this model may be more interesting for applications.
Similar content being viewed by others
References
Hopfield, J.J.: Proc. Nat. Acad. Sci. USA79, 2554 (1982)
Amit, D.J., Gutfreund, H., Sompolinsky, H.: Ann. Phys.173, 30 (1987)
Crisanti, Sompolinsky, H.: Phys. Rev. A36, 4922 (1987); Phys. Rev. A37, 4865 (1988)
Hertz, J.A., Thorbergsson, G.I.: Nordita-preprint 20, Kopenhagen (1988)
Kree, R., Zippelius, A.: Phys. Rev. A36, 4421 (1987)
Bös, S., Kühn, R., van Hemmen, J.L.: Z. Phys. B—Condensed Matter71, 261 (1988)
Virasoro, M.A.: Europhys. Lett.7, 293 (1988)
Horner, H.: Z. Phys. B—Condensed Matter (1989)
Krey, U., Pöppel, G.: Z. Phys. B—Condensed Matter76, 513 (1989)
Coolen, A.C.C., Denier van der Gon, J.J.: Preprint Thu-89-3, Utrecht (1989)
Personnaz, L., Guyon, I., Dreyfus, G.: Phys. Rev. A34, 4217 (1986)
Diederich, S., Opper, M.: Phys. Rev. Lett.58, 949 (1987)
Gardner, E.: J. Phys. A21, 271 (1988)
Krauth, W., Mezard, M.: J. Phys. A20 L745 (1987)
Pöppel, G., Krey, U.: Europhys. Lett.4 (1987)
Krey, U., Pöppel, G.: Proceedings of the Conference on Measures of Complexity, Rom, Oct. 1987. In: Lecture Notes in Physics. Peliti, L., Vulpiani, A. (eds.), Vol. 314, p. 54. Berlin, Heidelberg, New York: Springer 1988
Dehaene, S., Changeux, J.P., Nadal, J.P.: Proc. Natl. Acad. Sci. USA84, 2727 (1987)
Kühn, R., van Hemmen, J.L., Riedel, U.: J. Phys. A22, 3123 (1989)
Guyon, I., Personnaz, L., Nadal, J.P., Dreyfus, G.: Phys. Rev. A38, 6365 (1988)
Sompolinsky, H., Kanter, I.: Phys. Rev. Lett.57, 2861 (1986)
Gutfreund, H., Mezard, M.: Phys. Rev. Lett.61, 235 (1988)
Nadal, J.-P.: Proceedings of the Conference on Measures of Complexity, Rom, Oct. 1987. In: Lecture Notes in Physics. Peliti, L., Vulpiani, A. (eds.), Vol. 314, p. 35. Berlin, Heidelberg, New York: Springer 1988
Buhmann, J., Schulten, K.: Europhys. Lett.4, 1205 (1987)
Herz, A., Sulzer, R., Kühn, R., van Hemmen, J.L.: Biol. Cybern.60, 457 (1989)
Kerszberg, M., Zippelius, A.: Phys. Scr. (to be published)
Schmitz, H.-J., Pöppel, G., Wünsch, F., Krey, U.: J. Phys. (Paris)51, (1990)
If only a single time sequence σ(t) has to be stable, one may e.g. use the Hebbian prescriptionJ ij (t)=σ i (t)σ j (t−τ ij ), which has the same periodicity as σ(t)
For completion one allowss i =0 as input in the case of absent information
In the recognition of static patterns e.g. with the Hebb-rule learning, even a pure pattern is usually recognized with a small finite error probability, which is enhanced if the pattern resulting from one recognition process is used as an input to a subsequent recognition, similar to the dynamic case of sequences. This is not the case if the patterns are fixed points of the dynamics like in the case of the pseudo-inverse solution for the coupling matrix [11]
Engel, A., English, H., Schütte, A.: Europhys Lett.8(4), 393–397 (1989)
Herz, A.: In: Connectionism in perspective. Pfeiffer, R., Schreter, Z., Fogelman, F., Steels, L. (eds.). Amsterdam: Elsevier 1989
Shinomoto S.: Biol. Cybern.57, 197 (1987)
Author information
Authors and Affiliations
Rights and permissions
About this article
Cite this article
Bauer, K., Krey, U. On learning and recognition of temporal sequences of correlated patterns. Z. Physik B - Condensed Matter 79, 461–475 (1990). https://doi.org/10.1007/BF01437658
Received:
Revised:
Issue Date:
DOI: https://doi.org/10.1007/BF01437658