Emergent information from low-resolution representations. Methods and applications from protein simulations to artificial neural networks.

We live in an age characterised by informational abundance, in which not only scientific research but also contemporary society in general is shaped by the unprecedented production, availability, and diffusion of data. On the one hand, advances in measurement techniques, simulation methods, and computational power have enabled the description of natural systems at increasingly fine levels of detail and across broader spatial and temporal scales. On the other hand, the massive amount of digital data constitutes the essential fuel that feeds learning systems, making this abundance not only an epistemic resource for science but also a constitutive condition for the development of artificial intelligence. This thesis addresses this problem within a framework grounded in statistical physics and information theory, in which the configurational space provides the natural setting for the description of complex systems, and reduced representations -- namely simplified descriptions of a more detailed object of study -- are treated as tools for identifying the system properties most relevant to collective behaviour. From this perspective, the central questions that arise are how the informational content of a low-resolution description should be quantified; under what conditions informative reduced representations can be identified without supervision; and to what extent common organising principles can be recognised across systems of different microscopic nature, such as proteins and artificial neural networks. More generally, the aim of the present work is to clarify how reduced descriptions can provide not only computational simplification, but also conceptual access to the emergent organisation of complex systems.