Exhaustive Thermodynamical Analysis of Boolean Learning Networks

The learning and generalization capabilities exhibited by a recently introduced model of self-organizing Boolean networks are explained and interpreted by studying the thermodynamics of the training process of that model. The thermodynamical analysis shows that learning and generalization occur as a direct consequence of the second law of thermodynamics. The complexity of solving a problem, for a given architecture, can be precisely defined in terms of entropy changes and is related to the amount of specialization present in that architecture. We speculate our conclusions to be valid, to some extent, not only for our model, but also for more general and complex systems, including biological learning systems.