Deep learning has won great reputations over fields like computer vision, robotic, etc. In the field of physics, many efforts have also been paid to harness this power. In this talk, I'll introduce three such attempts of applying normalizing flow models to physics. Normalizing flow models are one kind of generative model that has an exact solvable probability, thus an ideal model for scientific usage. We will see this solvability can help us establish a mapping between a complex distribution and a simple one, and in turn, accelerate Monte Carlo sampling speed. By constructing a hierarchy structure, this mapping established by normalizing flows can be viewed as an information-preserving renormalization group (RG) transformation. This allows us to design RG mapping out of an auto-learning scheme. We can also build Hamiltonian mechanics into the network, the resulting normalizing flows can sort out collect variables according to their frequencies. This physics-informed model can be used to analyze complex dynamic processes like molecules.
Meeting Link: To request for meeting link, please write to firstname.lastname@example.org.