Periodic Reporting for period 1 - QuantAI (Artificial agency and learning in quantum environments)
Berichtszeitraum: 2022-10-01 bis 2025-03-31
Quantum mechanics is our most fundamental theory of physics. It has formed, and often challenged, our understanding of physical reality. We use quantum mechanics to manipulate and control matter and light at the atomic scale, and it provides the basis for many new technologies. At the same time, the rise of artificial intelligence (AI) and machine learning (ML) is gaining momentum in science and basic research. ML is already employed in different areas of physics, mostly for big data processing and classification. But the development of AI is heading much further and is likely to transform basic science in the near future.
In this project, we investigate the use of AI and artificial agency in basic science, with a focus on quantum physics and, more specifically, quantum information (QI). The overall objective of our project is to explore artificial agency, its scope and its limits, and what role it can play in quantum science. We want to develop models of artificial agency that are beneficial for basic research, both from a practical and a foundational perspective. Specific objectives include, among others, the following:
- Develop models of classical and quantum learning agents that can be fully integrated into a quantum environment. Use these models for the design of novel quantum experiments and QI applications, the exploration of quantum many-body states, and the investigation of near-term, quantum-enhanced AI technologies.
- Develop models of transparent learning agents, whose actions are interpretable by a human user. Develop these models towards learning agents that can be used for explorative experimenting and scientific discovery. Establish criteria for the attributability of agency to artificial entities that can be tested in experiments. Apply these criteria to AI applications in QI science and to foundational issues regarding the role of agents in quantum physics.
- Establish a physical framework for the discussion of interpretability, explainability and trustworthiness of quantum-enhanced AI. Develop and present classical models of learning agents that are explainable and that can be quantized.
The expected impact of project is both practical and fundamental. The integration of AI with physics at a fundamental level, including quantum physics, can have a dramatic and enduring impact on quantum science and beyond. It will generate many ideas for new protocols and experiments in QI, and it will help us to better control and understand the physics of quantum matter and complex systems. It will also provide a framework to discuss relevant physical aspects of questions regarding the interpretability, explainability and trustworthiness of quantum technologies and AI. This project will allow us to explore the scope and physical limits of artificial agents and what role they can play in basic research. Ultimately, this project will also help us to better understand—from a fundamental physics perspective—our role, as human agents, in science.
In this project, we investigate the use of AI and artificial agency in basic science, with a focus on quantum physics and, more specifically, quantum information (QI). The overall objective of our project is to explore artificial agency, its scope and its limits, and what role it can play in quantum science. We want to develop models of artificial agency that are beneficial for basic research, both from a practical and a foundational perspective. Specific objectives include, among others, the following:
- Develop models of classical and quantum learning agents that can be fully integrated into a quantum environment. Use these models for the design of novel quantum experiments and QI applications, the exploration of quantum many-body states, and the investigation of near-term, quantum-enhanced AI technologies.
- Develop models of transparent learning agents, whose actions are interpretable by a human user. Develop these models towards learning agents that can be used for explorative experimenting and scientific discovery. Establish criteria for the attributability of agency to artificial entities that can be tested in experiments. Apply these criteria to AI applications in QI science and to foundational issues regarding the role of agents in quantum physics.
- Establish a physical framework for the discussion of interpretability, explainability and trustworthiness of quantum-enhanced AI. Develop and present classical models of learning agents that are explainable and that can be quantized.
The expected impact of project is both practical and fundamental. The integration of AI with physics at a fundamental level, including quantum physics, can have a dramatic and enduring impact on quantum science and beyond. It will generate many ideas for new protocols and experiments in QI, and it will help us to better control and understand the physics of quantum matter and complex systems. It will also provide a framework to discuss relevant physical aspects of questions regarding the interpretability, explainability and trustworthiness of quantum technologies and AI. This project will allow us to explore the scope and physical limits of artificial agents and what role they can play in basic research. Ultimately, this project will also help us to better understand—from a fundamental physics perspective—our role, as human agents, in science.
The QuantAI project aims to advance artificial agency and learning within quantum environments. Throughout this reporting period, substantial progress was made towards the project's objectives.
Variational MBQC for Generative Quantum Modeling:
Measurement-Based Quantum Computation (MBQC) is a quantum computation model where measurements on entangled states drive the computational process. Typically, these measurements introduce randomness that must be corrected to perform accurate computations. By using the correction probability for quantum measurements as learning parameters, we enhanced the expressive capabilities of quantum circuits without adding quantum computational overhead. This achievement offers a unique framework for quantum generative tasks and paves the way for developing adaptable and effective quantum agents.
Projective Simulation (PS) and Hybrid AI Models:
Projective Simulation (PS) is a physical model of agency that utilizes an episodic and compositional memory (ECM) for decision-making and learning. We extended PS to the quantum domain by incorporating continuous time quantum walks, which can be implemented using photonic hardware. This quantum-enhanced version of PS can enable quantum agents to make decisions more efficiently, while also retaining a degree of interpretability—critical for understanding AI-driven research. This work highlighted an important trade-off between "quantumness" and traceability, providing new insights into the design of interpretable quantum learning agents.
AI-Driven Quantum Circuit Design:
We also explored the application of classical AI in quantum settings. Specifically, we developed transformer-based AI models capable of autonomously designing quantum circuits. Relying on advanced AI methods such as diffusion models, we succeeded in synthesizing circuits to produce quantum operations. This notable achievement was featured on the cover of Nature Machine Intelligence, underscoring its impact on advancing quantum technology through AI.
Interpretable Quantum Computational Gadgets:
Another major contribution was the creation of an automated pipeline for discovering reusable quantum computational subroutines—termed "gadgets." These gadgets, identified through machine learning techniques like reinforcement learning, sequence mining, and clustering, serve as building blocks for creating complex quantum circuits. This approach enables one to extract valuable, human-readable information from complex environments.
Variational MBQC for Generative Quantum Modeling:
Measurement-Based Quantum Computation (MBQC) is a quantum computation model where measurements on entangled states drive the computational process. Typically, these measurements introduce randomness that must be corrected to perform accurate computations. By using the correction probability for quantum measurements as learning parameters, we enhanced the expressive capabilities of quantum circuits without adding quantum computational overhead. This achievement offers a unique framework for quantum generative tasks and paves the way for developing adaptable and effective quantum agents.
Projective Simulation (PS) and Hybrid AI Models:
Projective Simulation (PS) is a physical model of agency that utilizes an episodic and compositional memory (ECM) for decision-making and learning. We extended PS to the quantum domain by incorporating continuous time quantum walks, which can be implemented using photonic hardware. This quantum-enhanced version of PS can enable quantum agents to make decisions more efficiently, while also retaining a degree of interpretability—critical for understanding AI-driven research. This work highlighted an important trade-off between "quantumness" and traceability, providing new insights into the design of interpretable quantum learning agents.
AI-Driven Quantum Circuit Design:
We also explored the application of classical AI in quantum settings. Specifically, we developed transformer-based AI models capable of autonomously designing quantum circuits. Relying on advanced AI methods such as diffusion models, we succeeded in synthesizing circuits to produce quantum operations. This notable achievement was featured on the cover of Nature Machine Intelligence, underscoring its impact on advancing quantum technology through AI.
Interpretable Quantum Computational Gadgets:
Another major contribution was the creation of an automated pipeline for discovering reusable quantum computational subroutines—termed "gadgets." These gadgets, identified through machine learning techniques like reinforcement learning, sequence mining, and clustering, serve as building blocks for creating complex quantum circuits. This approach enables one to extract valuable, human-readable information from complex environments.
The QuantAI project has led several advances that push the boundaries of both quantum computing and artificial intelligence, offering novel frameworks and tools.
Generative Quantum Modeling with MBQC:
A major achievement was the use of randomness in Measurement-Based Quantum Computation (MBQC) as a resource, rather than treating it as a nuisance.
Quantum Walks for Explainable AI:
We extended Projective Simulation (PS) into the quantum domain by incorporating quantum walks, creating a quantum variant of PS that highlights an interplay between the efficiency and interpretability of decision-making processes in the quantum domain. This work lays the foundation of future work that will explore this emergent trade-off to better understand the role of explainability in quantum AI.
AI-Driven Quantum Circuit Design:
By training transformer-based AI models to autonomously design quantum circuits, we introduced a powerful new tool for optimizing quantum computations. Future research will go into enhancing the transparency of these methods to ensure a reliable and broad adoption in quantum science.
Interpretable Quantum Computational Gadgets:
We developed a machine learning pipeline for the automated discovery of "gadgets," which are efficient and interpretable subroutines for quantum circuits.
Impact and Future Potential:
These advancements establish new methodologies for integrating AI with quantum technologies, and they open doors to a wide range of future applications in scientific exploration, optimization, and decision-making.
Generative Quantum Modeling with MBQC:
A major achievement was the use of randomness in Measurement-Based Quantum Computation (MBQC) as a resource, rather than treating it as a nuisance.
Quantum Walks for Explainable AI:
We extended Projective Simulation (PS) into the quantum domain by incorporating quantum walks, creating a quantum variant of PS that highlights an interplay between the efficiency and interpretability of decision-making processes in the quantum domain. This work lays the foundation of future work that will explore this emergent trade-off to better understand the role of explainability in quantum AI.
AI-Driven Quantum Circuit Design:
By training transformer-based AI models to autonomously design quantum circuits, we introduced a powerful new tool for optimizing quantum computations. Future research will go into enhancing the transparency of these methods to ensure a reliable and broad adoption in quantum science.
Interpretable Quantum Computational Gadgets:
We developed a machine learning pipeline for the automated discovery of "gadgets," which are efficient and interpretable subroutines for quantum circuits.
Impact and Future Potential:
These advancements establish new methodologies for integrating AI with quantum technologies, and they open doors to a wide range of future applications in scientific exploration, optimization, and decision-making.