Backpropagation: past and future

Abstract

Some scientists have concluded that backpropagation is a specialized method for pattern classification, of little relevance to broader problems, to parallel computing, or to our understanding of the human brain. The author questions these beliefs and proposes development of a general theory of intelligence in which backpropagation and comparisons to the brain play a central role. He also points to a series of intermediate steps and applications leading up to the construction of such generalized systems, including past applications to social science which in some ways go beyond the work in AI as such. The author presents a condensed mathematical summary of that work. He begins by summarizing a generalized formulation of backpropagation, and then discusses network architectures and applications which it opens up. >