Reliable learning algorithms are important for both fuzzy logic and neural networks. Widespread application of fuzzy logic can be significantly enhanced by robust and reliable methods for generating fuzzy logic rules from training data. Fuzzy logic learning must satisfy the principles of learning defined by neural net learning theory. The fundamental design and training issues for both fuzzy logic and neural net systems should be the same. These principles require the algorithm not only to design and train the net in polynomial time, but also to make an explicit attempt to generate the smallest possible net, which implies attempting to generate the smallest set of fuzzy rules to describe the phenomenon. That is, it implies making an explicit attempt to generalize. Generalization and polynomial time complexity are aspects of learning not often mentioned in the fuzzy logic literature. This paper is also an attempt to integrate and unify fuzzy logic and neural network architectures and learning. Since both fuzzy logic and neural networks arise from the same learning experience of humans, they must reside in the same computational structure in the human brain, not in different ones. One of the fundamental ideas of fuzzy systems is to better understand and explain complex decisions in natural, linguistic terms. It is shown here that this can be done using standard neural network architectures such as RBF nets.