Quantum Technology Initiative Journal Club

Thursday 7 May 2026, 16:00 → 17:00 Europe/Zurich

513/R-070 - Openlab Space (CERN)

Cenk Tüysüz (CERN), Michele Grossi (CERN)

Weekly Journal Club meetings organised in the framework of the CERN Quantum Technology Initiative (QTI) to present and discuss scientific papers in the field of quantum science and technology. The goal is to help researchers keep track of current findings and walk away with ideas for their own research. Some previous knowledge of quantum physics would be helpful, but is not required to follow the talks.

To propose a paper for discussion, contact: michele.grossi@cern.ch

16:00 → 17:00 CERN QTI Journal CLUB

Conveners: Cenk Tüysüz (CERN), Dr Michele Grossi (CERN)

16:00 Bridging expressivity and trainability in classical and quantum generative models

This talk is Maria's PhD defense rehearsal

Abstract:
Quantum computing provides a compelling framework for representing and sampling complex probability distributions, motivating the study of quantum generative models. Despite this conceptual appeal, existing approaches in quantum machine learning face fundamental obstacles related to trainability and scalability. In particular, highly expressive quantum models often suffer from challenges associated with gradient evaluation, limiting their practical applicability. The absence of a backpropagation-like training mechanism further contributes to these challenges.

This thesis investigates quantum energy-based models with the goal of improving their practical relevance for learning tasks. In this context, quantum Boltzmann machines are a prominent class of quantum machine learning models, offering a direct connection between statistical physics and machine learning. However, generic quantum Boltzmann machines formulations involve non-commuting operators that render gradient evaluation computationally demanding and restrict their usability.

To address these limitations, we introduce semi-quantum Boltzmann machines, and in particular semi-quantum restricted Boltzmann machines. By enforcing commutativity in the visible subspace while retaining non-commuting operators on hidden units, semi-quantum Boltzmann machines admit closed-form expressions for output probabilities and gradients. This constitutes a new class of quantum energy-based models that enable scalable training while preserving quantum features, and establish a direct correspondence with classical Boltzmann machines.

Building on this framework, we further develop a generalized contrastive divergence training algorithm for quantum generative models. The proposed method enables sample-based learning directly on quantum hardware, avoids explicit access to the full model distribution, and achieves constant-cost scaling per update step, analogous to local learning rules in classical energy-based models.

The theoretical results presented in this thesis are supported by numerical experiments with up to 100 qubits, which validate the analytical predictions and demonstrate the learning behavior of the proposed models. Together, these results provide a structured and trainable approach to quantum energy-based modeling and outline a scalable pathway for quantum generative learning on future fault-tolerant quantum hardware.

Speaker: Maria Demidik (DESY)

recording