Condensed Matter Seminar: Information bottlenecks and critical phenomena

Information-theory has long provided alternative viewpoints on fundamental physics: The notion of entropy as information, statistical mechanics as maximal entropy inference, and the characterization of RG flows based on their entanglement entropy. Following recent advancements in deep learning, a new information-theoretic quantity- the information bottleneck (IB), became numerically accessible but its correspondence with physics has been far less explored. In this talk, I’ll introduce IB and present several relations we established between IB and physical concepts such as renormalization, order parameters, and transfer matrices. Building on this I'll discuss potential ways in which IB, in conjugation with deep-learning, can shed light on outstanding problems in critical phenomena. In particular by providing an efficient numerical route for obtaining transfer matrix spectra and eigenvectors, even for large 3d problems.