Object Recognition by Composition

In order to realize machines that work in the complex environments of the “real world,” the most fundamental problem, I believe, is to devise a probabilistic model which can capture the nature of the world. Without such models, machines cannot learn, because there are too much to learn about the environment. Stuart Geman and his colleagues have developed a theory of “composition systems,” which is about to make hierarchical representations of objects in terms of parts and their relations. This might remind you of researches in syntactical pattern recognition well-studied in 1980’s. In the composition systems, the hierarchical representations are defined via composition rules which are syntactic rules, as production rules are, but come from the other direction. Since I feel the composition systems are very useful for real world problems, here I give a brief nontechnical view on the systems. In the paper entitled “Neural networks and the bias/variance dilemma,” they concluded that (see p.3 of [GBD92]) “learning complex tasks is essentially impossible without the a prior