ASSCOUNT: Associative Memory for Time Sequences

  • Berndt Müller
  • Joachim Reinhardt
Part of the Physics of Neural Networks book series (NEURAL NETWORKS)

Abstract

An obvious way to teach a Hopfield-type network to reproduce time sequences instead of stationary patterns is to introduce synaptic coefficients with an intrinsic time dependence [Am88, Gu88] as discussed in Sect. 3.4. Thus instead of the synaptic matrix w ij , which acts instantaneously to generate the local field defined in (20.1), one may introduce a collection of synapses w ij τ acting with a distribution of characteristic time delays of magnitude τ. The local field, which according to (20.2) determines the probability for the updated neuron state s i (t + 1), can be replaced by a generalized convolution in time
$$ {h_i}\left( t \right)\,\, = \,\,\sum\limits_{\tau \, = \,0}^{{\tau _{\max }}} {} \sum\limits_{j\, = \,1}^N {{\lambda ^\tau }w_{ij}^\tau {s_j}\left( {t\,\, - \,\,\tau } \right).} $$
(21.1)

Keywords

Convolution 

Preview

Unable to display preview. Download preview PDF.

Unable to display preview. Download preview PDF.

Copyright information

© Springer-Verlag Berlin Heidelberg 1990

Authors and Affiliations

  • Berndt Müller
    • 1
  • Joachim Reinhardt
    • 2
  1. 1.Department of PhysicsDuke UniversityDurhamUSA
  2. 2.Institut für Theoretische PhysikJ.-W.-Goethe-UniversitätFrankfurt 1Fed. Rep. of Germany

Personalised recommendations