Abstract

There is a great hope (and hype) that artificial neural networks will be able one day to solve all those cognitive tasks that the human brain is so accustomed to, but that today's machines find so difficult: perception, control, and reasoning. However, most models are facing an important 'scaling problem' when they try to cope with real-world tasks. Even the use of massively parallel machines or neural chips for the simulation of extremely large artificial neural networks does not guarantee the qualitative leap necessary to the solution of complex problems, which would finally bring artificial neural networks out of the research labs. One of the reasons that explain the partial failure of classical artificial neural models -- despite the considerable amount of funds and research teams involved -- can be found by looking back at the biological system from which the artificial models were inspired. While the great majority of neural network researchers still try to tackle complex problems by using larger and larger networks of uniform architecture, the human brain achieves its unmatched cognitive abilities by the collaboration of many, partially independent subsystems. Although we are far from understanding the pattern of connectivity and the processing stages of the human brain, let apart the arousal of symbolic reasoning from subsymbolic activity, a handful of researchers are pursuing a pioneering work aimed to the definition and exploration of modular architectures. This (mostly empirical) work has regularly shown the better adequacy of modular architectures to the solution of difficult tasks. Unfortunately, the lack of theoretical foundations leads to an unsatisfied need of simulation tools specifically aimed at the definition and study of modular systems, which considerably slows down the progress of the work and reduces the amount of research in the field. This thesis tries to give an answer to these problems, by proposing a mathematically sound framework -- the n-calculus -- suited to the specification of a broad class of modular systems, and a corresponding simulation environment that should enable the investigation of modular architectures with greater ease.

Citations