Adapting connectionist learning to Bayes networks

AbstractA clear need exists within artificial intelligence for flexible systems capable of modifying their own knowledge bases. A common formalism can be used to describe two seemingly different models: the Boltzmann machine connectionist learning model and the Bayes network model for probabilistic reasoning. The learning algorithm for Boltzmann machines can be adapted to a general algorithm for adjusting conditional probabilities on the links in a Bayes network. It is hypothesized that the formal approach outlined here holds promise for unifying symbolic and subsymbolic levels of reasoning