Semantic phase transition in a classifier based on an adaptive fuzzy system

摘要

We investigate how many rules a particular adaptive fuzzy system, very similar to a feedforward neural network, needs to perform digit classification. Both training of systems with different numbers of fuzzy rules and the application of a pruning technique to a non minimal system support the statement that one rule for each class is needed. In particular, a semantic phase transition is observed when a new rule is added to a system with 9 rules. This behaviour, which is not common for a feedforward neural network, could be ascribed to the derivation of the studied system from a fuzzy logic framework.