Abstract
Information processing in synchronous neural network models has
been of interest over some time. A replica symmetric analysis of
Little's model indicates the presence of period-two cycles in a
low-temperature paramagnetic phase for negative values of the
self-interaction between units 1. There has also been interest
in the dynamics and the stationary states of attractor neural
networks that process sequences of patterns. Fixed-point solutions
and cycles of period two were found recently for symmetric
sequence processing in a feed-forward layered network model 2.
We study here the synchronous dynamics of a recurrent neural
network with a Hebbian plus symmetric sequential learning rule.
The model has been of interest over some time in order to explain
experimental recordings in the cortex of monkeys. An exact
generating functional approach (GFA) 3 is used here, in the
pattern saturation limit, to obtain first the stationary
fixed-point solutions that characterize various equilibrium
phases. The further behavior of the network is then analyzed
through a numerical simulation procedure for the effective single
neuron dynamics obtained exactly using the GFA 4. The simulation
is first tested on the equilibrium results for a dominant Hebbian
term and the procedure is then used to look for cyclic solutions
that appear depending on the self-interaction $J_0$. For
$J_0>0$, only a fraction of spins flip at each time step, as in
Little's model. For $J_0<0$, a variety of period-two cyclic
solutions are obtained in this work with macroscopic parameters
that change sign at each time step, and characterize a retrieval,
a spin-glass or a high-$T$ paramagnetic phase. All spins flip at
every time step only for $T 0$ and sufficiently large
negative values of $J_0$. For $J_0=0$ and a dominant sequence
processing term, the network exhibits a cyclic behavior with
oscillating positive overlaps and spin-glass order parameter. The
results shown here provide a new insight into the nature of the
cyclic states in the parallel dynamics of recurrent neural
networks, which could be of interest for information processing.
Further features of sequence processing and open questions will be
pointed out.
1) J. F. Fontanari, J. Phys. France 49, 13 (1988). \\
2) F. L. Metz and W. K. Theumann, Phys. Rev. E 75, in press.\\
3) A. C. C. Coolen, in Handbook of Biological Physics IV: Neuro-Informatics and Neural Modeling, edited by F. Moss and S. Gielen (Elsevier, Amsterdam, 2001), p. 619.\\
4) H. Eissfeller and M. Opper, Phys. Rev. Lett. 68, 2094 (1992).\\
This work was partially supported by the Brazilian agencies CNPq and FAPERGS.
Links and resources
Tags