The combinatorial neural model (CNM) is a type of fuzzy neural network
for classification problems and, more generally, for the mapping betw
een fuzzy multidimensional spaces. Learning in CNM is a complex task s
panning the learning of input-neuron membership functions, of the netw
ork topology, and of connection weights. In this paper we are concerne
d with these various aspects of learning in CNM, most notably with the
learning of connection weights, whose complexity comes from the exist
ence of nondifferentiable, nonconvex error functions associated with t
he learning process. We introduce several algorithms for weight learni
ng, most based on subgradient techniques borrowed from the held of non
differentiable optimization. All algorithms are based on essentially '
'local'' rules, and are therefore amenable to distributed/parallel imp
lementations. Experimental results are provided on the large-scale pro
blem of monitoring the deforestation of the Amazon region on satellite
images. What these results indicate is that a hybrid CNM system outpe
rforms previous results obtained with variations of error backpropagat
ion techniques. In addition, this hybrid system has demonstrated robus
tness in the contest under consideration, therefore constituting an at
tractive alternative.