DEEP-INセミナー
28 イベント
-
セミナー
Neural Network Quantum States for Quarkonium in Medium: Real-Time Open Quantum-System Dynamics
2025年12月19日(金) 16:00 - 17:00
Tom Magorsch (Ph.D. Student, Department of Physics, Technical University of Munich, Germany)
Many phenomena in high energy physics can not be described by Euclidean-time Monte Carlo estimates alone, but require genuine real-time evolution and a treatment of non-equilibrium effects. However, such simulations are computationally challenging. One such example is the evolution of heavy quarkonium in the quark gluon plasma produced in heavy-ion collisions. In this talk, I will introduce the open quantum system treatment of in-medium quarkonium. I will then give an overview on neural network quantum states as a variational approach to the real-time simulation of open quantum systems. As a controlled benchmark system, I will study the application to the Caldeira-Leggett model and conclude with an outlook on future applications of neural network based simulation of quarkonia in medium.
会場: via Zoom
イベント公式言語: 英語
-
セミナー
A bi-fidelity Asymptotic-Preserving Neural Network approach for multiscale kinetic problems
2025年12月17日(水) 11:00 - 12:00
Liu Liu (Assistant Professor, Department of Mathematics, The Institute of Mathematical Sciences, The Chinese University of Hong Kong, Hong Kong)
In this talk, we will introduce a bi-fidelity Asymptotic-Preserving Neural Network (BI-APNNs) framework, designed to efficiently solve forward and inverse problems for the linear Boltzmann equation. Our approach builds upon the previously studied Asymptotic-Preserving Neural Network (APNNs), which employs a micro-macro decomposition to handle the model’s multiscale nature. We specifically address a bottleneck in the original APNNs: the slow convergence of the macroscopic density in the near fluid-dynamic regime. This strategy significantly accelerates the training convergence as well as improves the accuracy of the forward problem solution, particularly in the fluid-dynamic limit. We show several numerical experiments on both linear Boltzmann and the Boltzmann-Poisson system that this new BI-APNN method produces more accurate and robust results for forward and inverse problems compared to the standard APNNs. This is a joint work with Zhenyi Zhu and Xueyu Zhu.
会場: セミナー室 (359号室) (メイン会場) / via Zoom
イベント公式言語: 英語
-
セミナー
Generative sampling with physics-informed kernels
2025年12月8日(月) 14:00 - 15:00
Renzo Kapust (Ph.D. Student, Institute for Theoretical Physics, University Heidelberg, Germany)
We construct a generative network for Monte-Carlo sampling in lattice field theories and beyond, for which the learning of layerwise propagation is done and optimised independently on each layer. The architecture uses physics-informed renormalisation group flows that provide access to the layerwise propagation step from one layer to the next in terms of a diffusion equation for the respective renormalisation group kernel through a given layer. Thus, it transforms the generative task into that of solving once the set of independent and linear differential equations for the kernels of the transformation. As these equations are analytically known, the kernels can be refined iteratively. This allows us to structurally tackle out-of-domain problems generally encountered in generative models and opens the path to further optimisation. We illustrate the practical feasibility of the architecture within simulations in scalar field theories.
会場: セミナー室 (359号室) (メイン会場) / via Zoom
イベント公式言語: 英語
-
セミナー
Hamiltonian Learning and Dynamics Prediction via Machine Learning
2025年11月26日(水) 15:00 - 16:00
李 可仁 (Assistant Professor, College of Physics and Optoelectronic Engineering, Shenzhen University, China)
Accurate prediction of quantum Hamiltonian dynamics and identification of Hamiltonian parameters are crucial for advancements in quantum simulations, error correction, and control protocols. This talk introduces a machine learning model with dual capabilities: it can deduce time-dependent Hamiltonian parameters from observed changes in local observables within quantum many-body systems, and it can predict the evolution of these observables based on Hamiltonian parameters. The model’s validity was confirmed through theoretical simulations across various scenarios and further validated by two experiments. Initially, the model was applied to a Nuclear Magnetic Resonance quantum computer, where it accurately predicted the dynamics of local observables. The model was then tested on a superconducting quantum computer with initially unknown Hamiltonian parameters, successfully inferring them. We believe that machine learning techniques hold great promise for enhancing a wide range of quantum computing tasks, including parameter estimation, noise characterization, feedback control, and quantum control optimization.
会場: via Zoom
イベント公式言語: 英語
-
セミナー
On the Role of Hidden States of Modern Hopfield Network in Transformer
2025年11月10日(月) 14:00 - 15:00
瀧 雅人 (立教大学 大学院人工知能科学研究科 准教授)
Large language models such as ChatGPT are based on deep learning architectures known as Transformers. Owing to their remarkable performance and broad applicability, Transformers have become indispensable in modern AI development. However, it still remains an open question why Transformers perform so well and what the essential meaning of their unique structure is. One possible clue lies in the mathematical correspondence between Hopfield Networks and Transformers. In this talk, I will first introduce the major developments over the past decade that have significantly increased the storage capacity of Hopfield Networks. I will then review the theoretical correspondence between Hopfield Networks and Transformers. Building on this background, I will present our recent findings: by extending this correspondence to include the hidden-state dynamics of Hopfield Networks, we discovered a new class of Transformers that can recursively propagate attention-score information across layers. Furthermore, we found, both theoretically and experimentally, that this new Transformer architecture resolves the “rank collapse” problem often observed in conventional multi-layer attention. As a result, when applied to language generation and image recognition tasks, it achieves performance surpassing that of existing Transformer-based models.
会場: セミナー室 (359号室) (メイン会場) / via Zoom
イベント公式言語: 英語
-
セミナー
Quantum multi-body problems using unsupervised machine learning
2025年11月5日(水) 15:00 - 16:00
内藤 智也 (東京大学 大学院工学系研究科 原子核国際専攻 特任助教)
I will introduce the recent development of a method to calculate the (anti)symmetrized wave functions and energies of the ground and low-lying excited states using the unsupervised machine learning technique. I will also introduce the recent attempts to consider the spin-isospin degrees of freedom and extend them to the Dirac equation.
会場: セミナー室 (359号室) (メイン会場) / via Zoom
イベント公式言語: 英語
-
セミナー
Neural network wavefunctions for SU(2) lattice gauge theory in the Hamiltonian formulation
2025年10月22日(水) 15:00 - 16:30
Tom Spriggs (PostDoc, Kavli Institute of Nanoscience and QuTech, Delft University of Technology, Netherlands)
In this talk I will cover our recent preprint arXiv:2509.12323 where we propose a neural network approach to finding the ground state wavefunction of SU(2) lattice gauge theory. Specifically, we demonstrate that the use of bespoke SU(2)-gauge-equivariant neural network layers increases the extent to which our variational ansatz can represent the ground state of this system. During this talk I will contrast the Hamiltonian and Euclidean formalisms of lattice gauge theories, highlighting the promises that the former offers but also the difficulties: noting briefly the issues of parameterising the continuous Hilbert space that plague tensor network and quantum simulation approaches and how our approach alleviates this. I will try and present our method pedagogically as we are very interested in learning its uses but also the limits of its validity, before closing with some remarks on scaling to larger systems and different gauge groups.
会場: via Zoom
イベント公式言語: 英語
-
セミナー
Discovering and harnessing symmetry with machine learning
2025年10月6日(月) 16:00 - 17:30
Escriche Santos Eduardo (Ph.D. Student, Department of Computer Science, Technical University of Munich, Germany)
Incorporating symmetry-inspired inductive biases into machine learning models has led to many significant advances in the field, especially for its application to scientific data. However, recently, a trend has emerged that favors implicitly learning relevant symmetries from data instead of designing constrained equivariant architectures. In this talk, I will first introduce these different modelling alternatives, together with their associated benefits and limitations. Then, I will describe some examples of automatic symmetry discovery methods as a way of mitigating some of those limitations. Finally, I will present our recent work that integrates symmetry discovery and the definition of an equivariant model into a joint learnable end-to-end approach, which further alleviates some of the limitations of current equivariant modelling approaches.
会場: via Zoom
イベント公式言語: 英語
-
セミナー
Statistical Physics of In-Context Learning in Transformer
2025年9月16日(火) 15:00 - 16:30
Haiping Huang (Professor, School of Physics, Sun Yat-sen University, China)
The pre-trained large model demonstrates the ability to learn from examples, that is, it can infer patterns and generalize from a small number of examples without retraining. How does this ability emerge? This report proposes a physical model mapping of the large model pre-training process, and finds that the training process corresponds to spin condensation, the unique energy ground state will determine the example generalization ability, and the diversity of training data is a key element in algorithm design. This study also reveals that the reasoning process of the large model may be fundamentally different from human thinking.
会場: via Zoom
イベント公式言語: 英語
-
セミナー
Neural Network for Holographic QCD
2025年9月12日(金) 10:30 - 11:30
ホーンアン・ツオン (Ph.D. Candidate, College of Physics, Jilin University, China)
Holographic QCD provides a powerful theoretical framework for investigating the equation of state of boundary field theories, where the idea is that the boundary dynamics can be fully determined by solving the bulk equations of motion. However, the coupling functions in the action typically rely on external inputs (such as lattice QCD data), and their explicit forms are often based on artificial assumptions. To eliminate such arbitrariness, we introduce neural networks into the potential reconstruction framework to represent the coupling functions, thereby constructing a fully data-driven machine learning model governed solely by boundary field theory inputs. The results obtained after training show remarkable consistency with the coupling functions derived from holographic renormalization based on prior assumptions, highlighting the strong function-approximation capability of neural networks and revealing the potential to unify the potential reconstruction and holographic renormalization approaches within a common framework.
会場: セミナー室 (359号室) 3階 359号室 (メイン会場) / via Zoom
イベント公式言語: 英語
-
セミナー
Femtoscopy: Probing Fundamental Matter Properties at the Fermi Scale
2025年8月22日(金) 13:30 - 15:00
Yijie Wang (Postdoctoral Researcher, Tsinghua University, Beijin, China / 理化学研究所 仁科加速器科学研究センター (RNC) RI物理研究室 客員研究員)
Femtoscopy, a cutting-edge technique grounded in intensity interferometry (correlation function analysis), enables in-depth exploration of fundamental properties of matter, including space, time, and interactions, at the Fermi scale. Originating from the Hanbury Brown and Twiss (HBT) correlation function, which was utilized in 1956 to measure the angular radius of Sirius, this method has been extended to the subatomic realm, emerging as a pivotal tool for deciphering space-time structures and particle interactions. This talk focuses on three representative femtoscopy studies: First, by combining femtoscopic interferometry with optical deblurring algorithms, it reveals the non-Gaussian freeze-out spatial distribution of protons and antiprotons in Au+Au relativistic heavy-ion collisions, challenging conventional wisdom [Chinese Physics Letters 42, 031401 (2025)]. Second, in the 30 MeV/u ⁴⁰Ar + ¹⁹⁷Au reaction, femtoscopy is employed to determine the proton emission timescale at approximately 100 fm/c and uncover the kinetic law of preferential emission of neutron-rich particles, making an “ultra-fast” video for heavy-ion collisions [Physics Letters B, 825, 136856 (2022)]. Third, using a high-resolution neutron array, femtoscopy accurately measures the neutron-neutron scattering length and effective range, as well as the space-time size of the neutron emission source, providing crucial data for the study of charge symmetry breaking in nuclear forces and nuclear symmetry energy [Physical Review Letter, 134, 222301 (2025)]. These achievements fully demonstrate the significant value of femtoscopy in advancing the frontiers of nuclear and particle physics, spanning from experimental observations to theoretical modeling. Dr. Yijie Wang is the Post Doc of Tsinghua University and Visiting Scientist of RIKEN. He studied physics at Jilin University, China, and obtained his Ph. D. degree at Tsinghua University in 2021. Then, he continued researches in Tsinghua University as Post Doc up to now. His interests focus on heavy ion collision experiment, nuclear equation of state, advanced detection system development and femtoscopy.
会場: セミナー室 (359号室) 3階 359号室 (メイン会場) / via Zoom
イベント公式言語: 英語
-
セミナー
Generative Models for Statistical Field Theories
2025年6月25日(水) 15:00 - 16:00
王 凌霄 (理化学研究所 数理創造研究センター (iTHEMS) 数理基礎部門 研究員)
In the final talk of the DEEP-IN series, we will explore the role of generative models in learning phase transitions and sampling in lattice systems. First, we demonstrate how generative models can serve as global samplers by learning the underlying probability distributions. This enables the sampling of configurations more efficiently for lattice field theories. We will also demonstrate how the ferromagnetic phase transition, the Kosterlitz-Thouless transition, and quantum phase transitions can be identified from generative models. I will briefly introduce generative diffusion models, which can be interpreted as a stochastic quantization scheme. This opens a new path for understanding deep generative models. This is an informal seminar, we will start with the methodology and some practical examples, and finally reserve time for everyone interested to discuss it together.
会場: 理化学研究所 和光キャンパス 研究本館3階 345-347 (メイン会場) / via Zoom
イベント公式言語: 英語
-
セミナー
Gauge-equivariant multigrid neural networks
2025年6月19日(木) 10:30 - 12:00
Tilo Wettig (Professor, Universität Regensburg, Germany)
In lattice QCD simulations, the most time-consuming element is typically the solution of the Dirac equation in the presence of a given gauge field. The current state of the art is to use a multigrid preconditioner to reduce the condition number of the Dirac operator matrix. We show how such preconditioners can be constructed using gauge-equivariant neural networks. For the multigrid solve we employ parallel-transport convolution layers. For the multigrid setup we consider two versions: the standard construction based on the near-null space of the operator, and a gauge-equivariant construction using pooling and subsampling layers. We show that both versions eliminate critical slowing down.
会場: セミナー室 (359号室) 3階 359号室とZoomのハイブリッド開催
イベント公式言語: 英語
-
セミナー
Identifying Lightning Structures and Predicting Cloud Properties
2025年6月18日(水) 15:00 - 16:00
王 凌霄 (理化学研究所 数理創造研究センター (iTHEMS) 数理基礎部門 研究員)
This third talk in the DEEP-IN series focuses on using unsupervised machine learning to identify and predict patterns in atmospheric phenomena. We begin by demonstrating how clustering and dimensionality reduction techniques can uncover coherent lightning patterns from high-dimensional LOFAR (LOw Frequency ARray) data, offering insight into large-scale organization. We then show how generative diffusion models enable super-resolution retrieval of cloud properties for all day from satellite observations. This is an informal seminar, we will start with the methodology and some practical examples, and finally reserve time for everyone interested to discuss it together.
会場: セミナー室 (359号室) (メイン会場) / via Zoom
イベント公式言語: 英語
-
セミナー
Collective Behaviors and Deep Learning Applications
2025年6月11日(水) 15:00 - 16:00
王 凌霄 (理化学研究所 数理創造研究センター (iTHEMS) 数理基礎部門 研究員)
Understanding and modeling collective pedestrian behavior, particularly under extreme conditions, is a challenging problem that combines cognition, physics, and data analysis. In the second talk of DEEP-IN series, I will explore how deep learning can reveal the underlying principles of crowd dynamics from data. Starting with a bounded rationality framework, we demonstrate how deep learning can quantify evacuation dynamics and reveal hidden patterns in collective motion. Specifically, we demonstrate how macroscopic observables, such as entropy and kinetic energy, can be extracted from microscopic trajectories in simulations and real-world data. This is an informal seminar, we will start with the methodology and some practical examples, and finally reserve time for everyone interested to discuss it together.
会場: セミナー室 (359号室) (メイン会場) / via Zoom
イベント公式言語: 英語
-
セミナー
Solving Inverse Problems with Physics-Driven Deep Learning
2025年6月4日(水) 15:00 - 16:00
王 凌霄 (理化学研究所 数理創造研究センター (iTHEMS) 数理基礎部門 研究員)
This talk kicks off a four-part seminar series on the DEEP-IN WG, an interdisciplinary working group exploring how modern deep learning — including deep generative models — can tackle inverse problems across scientific domains. In addition to DEEP-IN activities, I will present a new framework and vision, motivated by the growing synergy between physics-driven designs for deep learning and scientific discovery, as discussed in our recent review article. Future talks will demonstrate machine learning applications in collective behaviors, weather systems, and lattice field simulations. This is an informal seminar, we will start with the methodology, give some practical examples, and finally reserve time for everyone interested to discuss it together.
会場: セミナー室 (359号室) (メイン会場) / via Zoom
イベント公式言語: 英語
-
セミナー
Can we infer probability distributions from cumulants? Probabilistic approaches to inverse problems
2025年3月18日(火) 15:30 - 16:30
Yang-Yang Tan (Ph.D. Candidate, Dalian University of Technology, China)
Inverse problems, which involve estimating system inputs from outputs, are prevalent across science and engineering. Their ill-posed nature often makes finding numerically stable and unique solutions challenging. This seminar explores probabilistic methods for reconstructing distributions from a finite set of their moments or cumulants. We apply the Maximum Entropy Method (MEM) and Gaussian Process (GP) to reconstruct net-baryon number distributions across the QCD chiral crossover region using cumulant data from the STAR experiment and functional renormalization group (fRG) calculations. Our results demonstrate how higher-order cumulants shape distribution tails, while anomalous features in the reconstructed distributions provide constraints on the input cumulants. We also discuss deep learning approaches for distribution reconstruction from cumulants and present our recent work on physics-informed neural networks (PINNs) for solving fRG equations.
会場: via Zoom
イベント公式言語: 英語
-
セミナー
Can AI understand Hamiltonian mechanics?
2025年1月31日(金) 16:00 - 17:00
Tae-Geun Kim (Ph.D. Student, Department of Physics, Yonsei University, Republic of Korea)
With recent breakthroughs in deep learning, particularly in areas like natural language processing and image recognition, AI has shown remarkable abilities in understanding complex patterns. This raises a fundamental question: Can AI grasp the core concepts of physics that govern the natural world? In this talk, as a first step towards addressing this question, we will discuss the possibility of AI understanding Hamiltonian mechanics. We will first introduce the concept of operator learning, a novel technique that allows AI to learn mappings between infinite-dimensional spaces, and its application to Hamiltonian mechanics by reformulating it within this framework. Then, we will test whether AI can derive trajectories in phase space given an arbitrary potential function, without relying on any equations or numerical solvers. We will then showcase our findings, demonstrating AI's capability to predict phase space trajectories under certain constraints. Finally, we will discuss the limitations, future research directions, and the potential for AI to contribute to scientific discovery.
会場: via Zoom
イベント公式言語: 英語
-
セミナー
Stochastic Normalizing Flows for Lattice Field Theory
2024年12月18日(水) 15:30 - 16:30
Elia Cellini (PhD, Department of Physics, University of Turin, Italy)
Normalizing Flows (NFs) are a class of deep generative models that have recently been proposed as efficient samplers for Lattice Field Theory. Although NFs have demonstrated impressive performance in toy models, their scalability to larger lattice volumes remains a significant challenge, limiting their application to state-of-the-art problems. A promising approach to overcoming these scaling limitations involves combining NFs with non-equilibrium Markov Chain Monte Carlo (NEMCMC) algorithms, resulting in Stochastic Normalizing Flows (SNFs). SNFs harness the scalability of MCMC samplers while preserving the expressiveness of NFs. In this seminar, I will introduce the concepts of NEMCMC and NFs, demonstrate their combination into SNFs, and outline their connections with non-equilibrium thermodynamics. I will conclude by discussing key aspects of SNFs through their application to Effective String Theory, SU(3) gauge theory, and conformal field theory.
会場: セミナー室 (359号室) 3階 359号室とZoomのハイブリッド開催
イベント公式言語: 英語
-
セミナー
How Neural Networks reduce the Fermionic Sign Problem and what we can learn from them
2024年12月11日(水) 15:30 - 16:30
Johann Ostmeyer (Post-doctoral Fellow, Helmholtz-Institut für Strahlen- und Kernphysik, University of Bonn, Germany)
When simulating fermionic quantum systems, non-perturbative Monte Carlo techniques are often the most efficient approach known to date. However, beyond half filling they suffer from the so-called sign problem, i.e. negative "probabilities", so that stochastic sampling becomes infeasible. Recently, considerable progress has been made in alleviating the sign problem by deforming the integration contour of the path integral into the complex plane and applying machine learning to find near-optimal alternative contours. In this talk, I am going to present a particularly successful architecture, based on complex-valued affine coupling layers. Furthermore, I will demonstrate how insight gained from the trained network can be used for simpler analytic approaches.
会場: via Zoom / セミナー室 (359号室) 3階 359号室とZoomのハイブリッド開催
イベント公式言語: 英語
28 イベント