D. Giataganas: "Stochastic Neural Networks as Thermodynamic Physical Systems"

Tuesday 15 Dec 2020, 14:00 → 16:00 Europe/Athens

Machine learning has been recently used as a very effective tool for the study and prediction of data in various fields of physics. Despite its enormous success, relatively little is understood theoretically about why these techniques are so successful. A possible starting point for a theoretical treatment suggests that deep learning is a form of coarse graining. On the other hand, in theoretical physics the notion of coarse graining is characterized by the renormalization group flow. We discuss the exciting possibility of an underlying fundamental relation between certain machine learning methods and the renormalization group flow in theoretical physics. To do so we will be introducing the restricted Boltzmann machines and their training methods on lattice spin models. While the stochastic neural networks have no direct knowledge about the Hamiltonian and the interactions of the physical models, we will show that they identify spontaneously the phase transitions of the lattice spin models with a process that turns out to resemble the renormalization group flow.

 

 

Join Zoom Meeting

https://us02web.zoom.us/j/85362733341

 

Dimokritos_15_12_2020.pdf