Chiara Cammarota

Rough-glassy landscapes from inference to machine learning

The evolution of many complex systems in physics, biology or computer science can often be thought of as an attempt to optimize a cost function. Such function generally depends on a highly non-linear way on the huge number of variables parametrizing the system so that its profile defines a high-dimensional landscape, which can be either smooth and convex, or rugged. In this talk I will focus on rough cost/loss functions within the realm of inference and machine learning. I will first discuss the cost landscape of a widely used inference model called spiked tensor model, hereby also generalised, and its implications on the performances of inference algorithms. Secondly I will report on evidences of glass-like dynamics, including aging, during training of deep neural networks, and use them to discuss the importance of over-parametrisation, widely used in the field.