R. Kozma et Wj. Freeman, Chaotic resonance - Methods and applications for robust classification of noisy and variable patterns, INT J B CH, 11(6), 2001, pp. 1607-1629
A fundamental tenet of the theory of deterministic chaos holds that infinit
esimal variation in the initial conditions of a network that is operating i
n the basin of a low-dimensional chaotic attractor causes the various traje
ctories to diverge from each other quickly. This "sensitivity to initial co
nditions" might seem to hold promise for signal detection, owing to an impl
ied capacity for distinguishing small differences in patterns. However, thi
s sensitivity is incompatible with pattern classification, because it ampli
fies irrelevant differences in incomplete patterns belonging to the same cl
ass, and it renders the network easily corrupted by noise. Here a theory of
stochastic chaos is developed, in which aperiodic outputs with 1/f(2) spec
tra are formed by the interaction of globally connected nodes that are indi
vidually governed by point attractors under perturbation by continuous whit
e noise. The interaction leads to a high-dimensional global chaotic attract
or that governs the entire array of nodes. An example is our spatially dist
ributed KIII network that is derived from studies of the olfactory system,
and that is stabilized by additive noise modeled on biological noise source
s. Systematic parameterization of the interaction strengths corresponding t
o synaptic gains among nodes representing excitatory and inhibitory neuron
populations enables the formation of a robust high-dimensional global chaot
ic attractor. Reinforcement learning from examples of patterns to be classi
fied using habituation and association creates lower dimensional local basi
ns, which form a global attractor landscape with one basin for each class.
Thereafter, presentation of incomplete examples of a test pattern leads to
confinement of the KIII network in the basin corresponding to that pattern,
which constitutes many-to-one generalization. The capture after learning i
s expressed by a stereotypical spatial pattern of amplitude modulation of a
chaotic carrier wave. Sensitivity to initial conditions is no longer an is
sue. Scaling of the additive noise as a parameter optimizes the classificat
ion of data sets in a manner that is comparable to stochastic resonance. Th
e local basins constitute dynamical memories that solve difficult problems
in classifying data sets that are not linearly separable. New local basins
can be added quickly from very few examples without loss of existing basins
. The attractor landscape enables the KIII set to provide an interface betw
een noisy, unconstrained environments and conventional pattern classifiers.
Examples given here of its robust performance include fault detection in s
mall machine parts and the classification of spatioternporal EEG patterns f
rom rabbits trained to discriminate visual stimuli.