Soutenance de thèse: Maxime Sevelev

Phase diagram, jamming and glass transitions in the non-convex perceptron

Maxime Sevelev

This thesis treats the «spherical perceptron model», a simple exactly solvable model for glassy behavior and jamming suitably generalized to negative values of scalar product parameter κ. The classical machine-learning problem of random pattern classification by the perceptron is a convex constraint satisfaction problem (CSP). Even when the «stability parameter» κ of the model becomes negative, the problem still makes sense and can be interpreted as the problem of particles on an N-dimensional sphere trying to avoid randomly placed obstacles. In this case, the corresponding CSP is non-convex. This thesis studies the problem in detail in the non-convex domain. Systematic study is made possible by assigning to a constraint satisfaction problem its corresponding optimization version endowed with a Hamiltonian function (cost function) quantifying the violations of the constraints, as a function of the system’s configuration. The connection between random CSP and glassy phenomenology in physics is well known and has been explored in detail for models with discrete variables. The presence of continuous variables in the (spherical) perceptron model enables us to unveil, in random CSP, the characteristic SAT/UNSAT transition where the system transits from the satisfiable regime (where the ground state has zero energy) to the unsatisfiable one (where the ground state energy is positive). This phase transition can also be interpreted as a jamming transition similar to the one that exhibit models with frictionless spheres. The simplicity of the considered model allows the exact determination of the zero temperature phase diagram as a function of the control parameters: the density of obstacles and their size. In the present thesis, the jamming transition thus identified is completely characterized and several glass phases of stable and marginal character are studied in detail.

Keywords: spin glass, phase transition, disordered system, jamming

Directeur de thèse: S. Franz

Jury: R. Mulet, F. Ricci.tersenghi, V. Lecomte, M. Tarzia, V. Terras, C. Texier.

Date/Time : 06/10/2017 - 14:00 - 17:00

Location : IPN-batiment 100, Auditoriurm

Upcoming seminars