Quantum neuromorphic computing

Danijela, Julien and Baptiste opening the new dilution refrigerator @ UMPhy.

Quantum neuromorphic computing implements neural networks on quantum physical systems. Such physical quantum neural networks take advantage of the highly dimensional Hilbert space which allows to separate it in different classes. Moreover, quantum neural networks have the capacity to process input quantum states and learn to automatically recognize them, thus circumventing the need for a large number of classical measurements. In this perspective article: APL 2020 (arXiv), we review the possible different approaches to quantum neuromorphic computing.

In our team, we study quantum properties and dynamics in superconducting circuits in order to develop and implement new learning methods with the quantum neural networks. On this topic we collaborate closely with QCMX team at Ecole Polytechnique.

We are currently looking for a post-doc to work with us on experimental implementation of quantum neural networks, contact us if you are interested!

Everton, Danijela and Baptiste mounting a magnet on a dilution cryostat.
Fabrication of the superconducting circuit that will implement a quantum neural network.

Here are recent results that we have obtained:

Quantum Reservoir Computing: Reservoir computing is a machine learning paradigm that leverages the complex dynamics of a fixed, nonlinear system to transform input data into a high-dimensional representation, enabling efficient learning with minimal training. In our study, we implemented quantum reservoir computing using coherently coupled quantum oscillators, offering an alternative to qubit-based quantum neural networks. By leveraging the infinite basis states of bosonic modes, we created a densely connected quantum reservoir with up to 81 neurons using only two oscillators, significantly reducing hardware requirements compared to classical and qubit-based approaches. We demonstrated state-of-the-art accuracy of 99% on benchmark tasks such as sine-square waveform classification and chaotic time-series prediction, which would typically require at least 24 classical oscillators. Our analysis revealed that strong coupling and dissipation rates play crucial roles in reservoir performance, and that quantum coherence enhances computational efficiency. These findings highlight the scalability of our approach, suggesting that a system with just 10 oscillators could achieve billions of neurons, making quantum reservoir computing a promising platform for complex machine learning tasks (npj Quantum Information 2023).

Training bosonic systems with backprop: We developed an analog bosonic quantum neural network (QNN) that utilizes Fock basis measurements to train parametric interactions. Unlike traditional qubit-based quantum networks, our approach leverages bosonic modes, enabling data encoding and processing through trainable parametric drives with tunable amplitudes, phases, and frequency detunings. We implemented physics-aware training and backpropagation to optimize these parameters, improving performance on benchmark tasks such as waveform classification and image recognition. Our results show that training the bosonic QNN significantly reduces the number of required measurements compared to quantum reservoir computing while maintaining high accuracy. By demonstrating competitive results with fewer trainable parameters than classical and quantum alternatives, we highlight the potential of our approach for scalable quantum machine learning [arXiv].