DEEP-IN Seminar
5 events
-
SeminarUnderstanding Diffusion Models by Feynman's Path Integral
October 9 (Wed) at 15:00 - 16:30, 2024
Yuji Hirono (Assistant Professor, Department of Physics, Graduate School of Science, Osaka University)
Diffusion models have emerged as powerful tools in generative modeling, especially in image generation tasks. In this talk, we introduce a novel perspective by formulating diffusion models using the path integral method introduced by Feynman for describing quantum mechanics. We find this formulation providing comprehensive descriptions of score-based diffusion generative models, such as the derivation of backward stochastic differential equations and loss functions for optimization. The formulation accommodates an interpolating parameter connecting stochastic and deterministic sampling schemes, and this parameter can be identified as a counterpart of Planck's constant in quantum physics. This analogy enables us to apply the Wentzel-Kramers-Brillouin (WKB) expansion, a well-established technique in quantum physics, for evaluating the negative log-likelihood to assess the performance disparity between stochastic and deterministic sampling schemes.
Venue: Seminar Room #359 (Main Venue) / via Zoom
Event Official Language: English
-
Seminar
Renormalization Group Approach for Machine Learning Hamiltonian
September 10 (Tue) at 15:00 - 17:00, 2024
Misaki Ozawa (CNRS Researcher, Laboratory for Interdisciplinary Physics (LIPhy), Université Grenoble Alpes, France)
We develop a multiscale approach to estimate high-dimensional probability distributions. Our approach applies to cases in which the energy function (or Hamiltonian) is not known from the start. Using data acquired from experiments or simulations we can estimate the underlying probability distribution and the associated energy function. Our method—the wavelet-conditional renormalization group (WCRG)—proceeds scale by scale, estimating models for the conditional probabilities of “fast degrees of freedom” conditioned by coarse-grained fields, which allows for fast sampling of many-body systems in various domains, from statistical physics to cosmology. Our method completely avoids the “critical slowing-down” of direct estimation and sampling algorithms. This is explained theoretically by combining results from RG and wavelet theories, and verified numerically for the Gaussian and φ4-field theories, as well as weak-gravitational-lensing fields in cosmology. Misaki Ozawa obtained his Ph.D. in 2015 from the University of Tsukuba. He did his first postdoc at the University of Montpellier in France. He then moved to Ecole Normale Supérieure (ENS) Paris as the second postdoc. Currently, he is a CNRS permanent researcher at Grenoble Alpes Univeristy in France. His background is in the physics of disordered systems such as glasses and spin glasses. He is also working on interdisciplinary studies between statistical physics and machine learning.
Venue: #359, 3F, Seminar Room #359 / via Zoom
Event Official Language: English
-
Seminar
Symmetries and Generalization for Machine Learning on a Lattice
July 23 (Tue) at 15:00 - 16:30, 2024
Andreas Ipp (Senior Scientist, Institute for Theoretical Physics, TU Wien, Austria)
Symmetries such as translations and rotations are crucial in physics and machine learning. The global symmetry of translations leads to convolutional neural networks (CNNs), while the much larger space of local gauge symmetry has driven us to develop lattice gauge equivariant convolutional neural networks (L-CNNs). This talk will discuss how the challenges of simulating the earliest stage of heavy ion collisions led us to use machine learning and how these innovations could improve lattice simulations in the future. Andreas Ipp is a Senior Scientist at the Institute for Theoretical Physics at TU Wien. He received his PhD in 2003 and held postdoctoral positions at ECT* in Trento and the Max-Planck-Institute in Heidelberg before returning to TU Wien in 2009. He completed his habilitation on "Yoctosecond dynamics of the quark-gluon plasma" in 2014. His current research focuses on symmetries in machine learning for applications in lattice gauge theory and heavy ion collisions.
Venue: Seminar Room #359 (Main Venue) / via Zoom
Event Official Language: English
-
Seminar
Discovering Physical Laws with Artificial Intelligence
July 12 (Fri) at 10:00 - 11:30, 2024
Liu Ziming (Ph.D. Student, Department of Physics, Massachusetts Institute of Technology, USA)
Deep neural networks have been extremely successful in language and vision tasks. However, their black-box nature makes them undesirable for scientific tasks. In this talk, I will show how we can make these black-box AI models more interpretable and transparent and use them to discover physical laws, including conservation laws (AI Poincare), symmetries, phase transitions and symbolic relations (Kolmogorov-Arnold Networks). Ziming is a physicist and a machine learning researcher. Ziming received BS in physics from Peking Univeristy in 2020, and is current a fourth-year PhD student at MIT and IAIFI, advised by Max Tegmark. His research interests lie generally in the intersection of artificial intelligence (AI) and physics (science in general).
Venue: via Zoom
Event Official Language: English
-
Seminar
Inferring collective behavior from social interactions to population coding
June 27 (Thu) at 16:00 - 17:30, 2024
Chen Xiaowen (Postdoctoral Researcher, Laboratoire de Physique de l’École normale supérieure, CNRS, France)
(This is a joint iTHEMS Biology Seminar) From social animals to neuronal networks, collective behavior is ubiquitous in living systems. How are these behaviors encoded in interactions, and how do they drive biological functions? Recent insights from statistical physics applied to biological data have offer exciting new perspectives. However, previous research has mostly focused on the statics, i.e. the steady-state distributions of the collective behavior, without taking into consideration of time. In this talk, I will present two recent progresses tapping into the temporal domain. First, I will present a study of collective behavior in social mice from their co-localization patterns. To capture both static and dynamic features of the data, we developed a novel inference method termed the generalized Glauber dynamics (GGD) that can tune the dynamics while keeping the steady state distribution fixed. I will first outline the explanation power of the GGD dynamics, then explain how to infer the dynamics from data. The inferred interactions characterize sociability for different mice strains. In the second example, we studied information flow among neurons in the larval zebrafish hindbrain. By adapting the method of Granger causality to single cell calcium transient data, we were able to detect both a global information flow among neurons, as well as identifying brain regions that are key in locomotion.
Venue: via Zoom
Event Official Language: English
5 events