Workshop on Tensor Networks and (Quantum) Machine Learning for High-Energy Physics

4 Nov 2024, 09:00 → 5 Nov 2024, 17:30 Europe/Zurich

4/3-006 - TH Conference Room (CERN)

The workshop will be held on November 4-5, 2024, at CERN, Geneva, Switzerland. This event is organized as part of Task 1.4 within the Next Generation Triggers project, a new initiative that recently began its five-year journey. More details about the project can be found here: https://nextgentriggers.web.cern.ch/t14/.

The goal of the workshop is to bring together leading researchers and experts to explore cutting-edge methodologies and innovative applications at the intersection of Tensor Networks, Quantum Machine Learning, and High-Energy Physics.

The topics covered will include:

• Exploring the use of (Quantum) Machine Learning algorithms within tensor network wavefunctions.
• Analyzing the application of GPU technology for tensor network simulations in high-energy physics.
• Investigating new strategies for enhancing quantum machine learning using tensor networks.

 

Zoom connection will be available (see link from agenda)

Organising committee:

Stefano Carrazza

Enrique Rico Ortega

Sofia Vallecorsa

Michelangelo Mangano

Monday 4 November

09:00 → 13:05 Applications of Tensor Networks (TN) and Quantum Machine Learning (QML) to High-Energy Physics

09:30 Welcome coffee

10:00 Welcome and introduction

Speaker: Michelangelo Mangano (CERN)

Mangano_TN_Wshop_Intro.pdf

10:05 Quantum Machine Learning Integration in the High Energy Physics Pipeline

The integration of Quantum Machine Learning (QML) into the High Energy Physics (HEP) pipeline represents a transformative approach to addressing computational challenges in the analysis of vast and complex datasets. This talk will explore the synergy between quantum computing and machine learning, demonstrating how QML algorithms can handle HEP data sets and where QML has been applied to HEP challenges, such as anomaly detection, event classification, cross section integration and will outline future directions for incorporating quantum technologies into the broader HEP research framework and beyond.

Speaker: Dr Michele Grossi (CERN)

GrossiCERN_TNWorkshop24.pdf

11:05 TNS and the simulation of Lattice Gauge Theories

Tensor networks have demonstrated their suitability to describe equilibrium states of LGT in small spatial dimensions, where it has been possible to realize continuum limit extrapolations with record precision. And they are also a most adequate tool to design and benchmark quantum simulation protocols.
Real time evolution poses a more challenging scenario, which escapes the reach of the most traditional LGT methods but is physically relevant for experimental and fundamental questions. TNS methods can in principle address some dynamical problems, also for LGT, but are subject to limitations. Understanding them is furthermore crucial to determine a possible quantum advantage in these setups.
In this talk I will review the status of TNS simulations for LGT problems, with especial interest in recent and ongoing studies that tackle dynamical scenarios.

Speaker: Mari Carmen Bañuls

TNSforLGTdyn_MCBanuls_Nov24.pdf

12:05 Discussion: Utilizing QML Algorithms within Tensor Networks

13:00 → 15:00 Lunch break

15:00 → 18:30 Tensor Networks and Neural Networks

15:00 Boosting applications in nuclear physics via recent advances in tensor network state methods

Recently we have proposed combination of the valence-space in-medium similarity renormalization group (VS-IMSRG) with the density matrix renormalization group (DMRG) offering a scalable and flexible many-body approach for strongly correlated open-shell nuclei. Combined with an analysis of quantum information measures, this further establishes the VS-DMRG as a valuable method for ab initio calculations of nuclei. Here we briefly overview recent advances in tensor network state methods that have the potential to further boost the application of DMRG in nuclear physics. These include a general approach to find an optimal representation of a quantum many body wave function, i.e., a parametrization with the minimum number of parameters for a given error margin via global fermionic mode optimization, combination of DMRG with restricted active space method (DMRG-RAS-X), multi-orbital correlations and entanglement, developments on hybrid CPU-multiGPU parallelization, and an efficient treatment of non-Abelian symmetries on high performance computing (HPC) infrastructures. Scaling analysis on NVIDIA DGX-H100 platform is also presented, advertising that quarter petaflops performance can be reached on a single node.

Speaker: Ors Legeza

cern2024_v01.pdf

16:00 Coffee break

16:30 Quantum Machine Learning: A Computer Science and Engineering Perspective

Quantum Machine Learning is a promising new field that joins quantum computing and machine learning to obtain computational advantages in learning from data. As such, there are plenty of opportunities but also challenges in the short and medium term. In this talk, we will address some of them from the point of view of computer science and engineering, with examples and lessons learned from real projects in collaboration with different universities and research institutes.

Speaker: Elias Fernandez-Combarro

QML.pdf

17:30 Discussion: Accelerating Calculations with Specialized Hardware (TPUs, GPUs) for QML and TN

18:30 → 20:00 Welcome drink

Tuesday 5 November

09:00 → 13:20 Prospects on Tensor Networks and Machine Learning

09:00 Tensor network algorithms for HEP quantum simulation

We review some recent results on the development of efficient tree tensor network algorithms and their applications to quantum simulation benchmarking and theoretical interpretation. In particular, we present results on lattice gauge theories 2+1 and 3+1 dimensions at finite density, and out-of-equilibrium in 1+1 dimensions. Moreover, we present a roadmap for future tensor networks simulations of increasing complexity. Finally, we present the application of tensor network methods to the solution of hard classical combinatorial problems via mapping to many-body quantum hamiltonians and of tensor network machine learning for b-bbar tagging at LHCb.

Speaker: Simone Montangero (Padova University)

09:45 Optimising Tree Tensor Networks for classification on hardware accelerators

This talk presents the development of a framework for TTN based binary classification. The primary objective of this study is to train and evaluate the performance of TTN classifiers and optimise the code for their efficient deployment on computing accelerators, such as General Purpose Graphics Processing Units (GPGPU) or Field-Programmable Gate-Arrays (FPGA). To evaluate the effectiveness of the implementation, we show different TTN classifier models, trained on synthetic ML datasets as well as on physics data for more challenging classification tasks. In the context of HEP applications, the computational burden of these models becomes pivotal, therefore we discuss information-aware pruning methods based on the explainability feature of Quantum-inspired machine learning models. Moreover, simulating the hardware logic we evaluate other compression possibilities. In conclusion possible further developments of this software and its integration in more robust frameworks such as Quantum TEA are discussed.

Speaker: Alberto Coppi

Coppi-TTNs_on_hardware_accelerators.pdf

10:05 Tree Tensor Network implementation on FPGA

Starting from the statements of the previous talk, we present the implementation on FPGA of Tree Tensor Networks as binary classifiers, highlighting the possibility of their deployment in high-frequency real-time environments, such as the online trigger systems of HEP experiments. The linear algebra operations needed by TTNs make them easily deployable on FPGAs, which are extremely suitable for concurrent tasks like matrix multiplications and tensor contractions. In this talk, we provide an intuitive description of the inference firmware achieving sub-microsecond latency, together with a projection of the necessary hardware resources for different combinations of TTN hyperparameters and degrees of parallelization. We probe the feasibility of the hardware implementation of TN classifiers for HEP applications, exploring future prospects and further developments, possibly exploiting the usage of AI Engines on AMD Versal boards.

Speaker: Lorenzo Borella (Universita e INFN, Padova (IT))

TTNonFPGA.pdf

10:30 Morning coffee

11:00 Provable exponential quantum advantages in learning from physics data

One of the key challenges of the quantum machine learning field is identifying learning problems where quantum learning algorithms can achieve a provable exponential advantage over classical learning algorithms. Previous examples of provable advantages are all arguably contrived, and all rely on cryptographic methods to make learning hard for a classical learner. Further, they are not aligned with the general intuition that the first advantages should come in the learning of quantum systems such as encountered in high energy experiments. In this talk we show that this general intuition is nonetheless correct.
In the first part of this talk I will discuss new observations which allow us to make formal proofs of quantum learning advantages in a broad scope of settings, relying on widely believed conjectures in complexity theory. In the second part of the talk, I will how this can be applied to meaningful scenarios such the learning of unknown observables with sampling errors, with a provable advantage. We will also reflect on the possible consequences of these results on learning advantages in high energy scenarios

Speaker: Vedran Dunjko

CERN-2024-QML.pdf

12:00 Discussion: Using Tensor Networks to Enhance QML Algorithms