K. Pakdaman et al., TRANSIENT OSCILLATIONS IN CONTINUOUS-TIME EXCITATORY RING NEURAL NETWORKS WITH DELAY, Physical review. E, Statistical physics, plasmas, fluids, and related interdisciplinary topics, 55(3), 1997, pp. 3234-3248
A ring neural network is a closed chain in which each unit is connecte
d unidirectionally to the next one. Numerical investigations indicate
that continuous-time excitatory ring networks composed of graded-respo
nse units can generate oscillations when interunit transmission is del
ayed. These oscillations appear for a wide range of initial conditions
. The mechanisms underlying the generation of such patterns of activit
y are studied. The analysis of the asymptotic behavior of the system s
hows that (i) trajectories of most initial conditions tend to stable e
quilibria, (ii) undamped oscillations are unstable, and can only exist
in a narrow region forming the boundary between the basins of attract
ion of the stable equilibria. Therefore the analysis of the asymptotic
behavior of the system is not sufficient to explain the oscillations
observed numerically when interunit transmission is delayed. This anal
ysis corroborates the hypothesis that the oscillations are transient.
In fact, it is shown that the transient behavior of the system with de
lay follows that of the corresponding discrete-time excitatory ring ne
twork. The latter displays infinitely many nonconstant periodic oscill
ations that transiently attract the trajectories of the network with d
elay, leading to long-lasting transient oscillations. The duration of
these oscillations increases exponentially with the inverse of the cha
racteristic charge-discharge time of the neurons, indicating that they
can outlast observation windows in numerical investigations. Therefor
e, for practical applications, these transients cannot be distinguishe
d from stationary oscillations. It is argued that understanding the tr
ansient behavior of neural network models is an important. complement
to the analysis of their asymptotic behavior, since both living nervou
s systems and artificial neural networks may operate in changing envir
onments where long-lasting transients are functionally indistinguishab
le from asymptotic regimes.