Periodic Reporting for period 1 - qDynnet (Quantum dynamical neural networks)
Periodo di rendicontazione: 2023-03-01 al 2025-08-31
Quantum neural networks aim to harness uniquely quantum features—such as superposition and entanglement—to process and transform input data. They hold particular promise for tasks like classification and automatic recognition of quantum states, to which they can naturally couple. This capability could offer a powerful alternative to classical quantum state tomography, potentially solving the challenging problem of reconstructing quantum states from classical measurements.
The most common approach to building quantum neural networks is through parameterized quantum circuits based on qubits. These have already shown the potential to represent an exponential number of neurons and to outperform classical neural networks in learning and reconstructing quantum processes. However, they still face significant challenges related to scalability and trainability.
The qDynnet project proposes a fundamentally different architecture, based on parametrically coupled quantum oscillators. This shift replaces the traditional notion of physical connectivity with spectral connectivity: a single nonlinear physical connection can support multiple parametric interactions when driven at different frequencies. Each of these interactions is trainable via its amplitude, phase, and detuning, enabling dense connectivity and a large number of tunable parameters without increasing hardware complexity.
Inspired by computational neuroscience and neuromorphic computing, qDynnet seeks to compute with a dynamical, analog quantum system—an approach more closely aligned with the way the brain processes information.
The most common approach to building quantum neural networks is through parameterized quantum circuits based on qubits. These have already shown the potential to represent an exponential number of neurons and to outperform classical neural networks in learning and reconstructing quantum processes. However, they still face significant challenges related to scalability and trainability.
The qDynnet project proposes a fundamentally different architecture, based on parametrically coupled quantum oscillators. This shift replaces the traditional notion of physical connectivity with spectral connectivity: a single nonlinear physical connection can support multiple parametric interactions when driven at different frequencies. Each of these interactions is trainable via its amplitude, phase, and detuning, enabling dense connectivity and a large number of tunable parameters without increasing hardware complexity.
Inspired by computational neuroscience and neuromorphic computing, qDynnet seeks to compute with a dynamical, analog quantum system—an approach more closely aligned with the way the brain processes information.
Over the first 24 months of the project, we explored how quantum systems can process and transform classical data. Using the framework of quantum reservoir computing, we implemented an experimental setup with a superconducting resonator and a transmon qubit. We have shown that by encoding the classical information in the amplitude of a quantum state, and measuring it in the photon number basis, we achieve data expansion, which allows us to perform classification tasks. We also found that additional Kerr nonlinearity further improves classification performance. To account for the effects of this nonlinearity, we developed new methods for encoding the input data more effectively.
In parallel, our simulations revealed that quantum properties like coherence play a direct role in improving classification accuracy—when compared to a classical system of the equivalent size. Furthermore, we developed a new approach to training our quantum system. By combining machine learning techniques like backpropagation with concepts from gaussian boson sampling, we were able to fine-tune the parameters that govern how different parts of the system interact. With this approach, we successfully demonstrated learning on a widely used dataset of handwritten digits. Notably, the learning performance we achieved with six coupled quantum modes is impossible with only data expansion using the same hardware.
In parallel, our simulations revealed that quantum properties like coherence play a direct role in improving classification accuracy—when compared to a classical system of the equivalent size. Furthermore, we developed a new approach to training our quantum system. By combining machine learning techniques like backpropagation with concepts from gaussian boson sampling, we were able to fine-tune the parameters that govern how different parts of the system interact. With this approach, we successfully demonstrated learning on a widely used dataset of handwritten digits. Notably, the learning performance we achieved with six coupled quantum modes is impossible with only data expansion using the same hardware.
We have shown that by training the parameters directly within the quantum system, we can significantly reduce the number of measurements needed, compared to untrained approaches like quantum reservoir computing. In practical terms, this means that the number of measurements required for inference can be reduced by about a factor of 10. This improvement is crucial, as it greatly speeds up the process of reading out information from the system. Looking ahead, this has important implications for practical use of quantum neural networks, for example integrated with quantum computers for automatic recognition of their output quantum states.