A unified formalism for neural net training algorithms

The authors present a framework which provides both a unified formalism for describing many connectionist algorithms and a formal definition of the goal of learning for these algorithms. This formal approach is illustrated through several examples from among the classical connectionist literature. Many nonconnectionist systems also fall into this formulation which is thus very general and has several consequences on the design of connectionist systems. For example it allows the training of optimally hybrid architectures where different connectionist or classical modules interact