A network of fuzzy neurons is different from a traditional network because the function of each neuron is identified and its semantics is defined. The function of such networks is the modeling of inference rules for classification.

The outputs give a measurement of the realization of a rule, i.e. the membership of an expected class. If the class is described by only one rule, the knowledge can be represented as follows:

IF [( is ) ( is )] ... [( is ) ( is )] THEN (Y is B)

The operators are operators of conjunction or disjunction used to combine the premise. The structure is generally split into functional layers:

• the layer of input weights the values of input according to their importance,
• the hidden layers combine the premises and
• the layer of output measures the degree of adequacy of the premises for the class considered.
  

In such networks, the learning cannot be carried out by the traditional algorithm of backpropagation. The backpropagation must adapt the computation of the errors to the structure of each neuron.

The thesis of Mascarilla [Mas96] presents a study of various algorithms for the learning of fuzzy neural systems. The aim is the automatic extraction of knowledge for the interpretation of satellite images as in the ICARE system, and the associated rules to process these knowledge. The proposed system generates, from sample indicated by the user,

• a base of fuzzy rules with membership functions and optimized logical connectors,
• the linguistic interpretation of these rules in natural language,
• the set of neural network build according to these rules.