Handbook on Neural Information Processing

This handbook is inspired by two fundamental questions: ”Can intelligent learning machines be built?” and ”Can they be applied to face problems otherwise unsolv- able?”. A simple unique answer certainly does not exist to the first question. Instead in the last three decades, the great amount of research in machine learning has suc- ceeded in answering many related, but far more specific, questions. In other words, many automatic tools able to learn in particular environments have been proposed. They do not show an ”intelligent” behavior, in the human sense of the term, but certainly they can help in addressing problems that involve a deep perceptual un- derstanding of such environments. Therefore, the answer to the second question is also partial and not fully satisfactory, even if a lot of challenging problems (compu- tationally too hard to be faced in the classic algorithmic framework) can actually be tackled with machine learning techniques. In this view, the handbook collects both well-established and new models in con- nectionism, together with their learning paradigms, and proposes a deep inspection of theoretical properties and advanced applications using a plain language, partic- ularly tailored to non experts. Not pretending to be exhaustive, this chapter and the whole book delineate an evolving picture of connectionism, in which neural information systems are moving towards approaches that try to keep most of the information unaltered and to specialize themselves, sometimes based on biological inspiration, to cope expertly with difficult real–world applications.