Experimental neuromorphic computing based on quantum memristor
Mirela Selimović, Iris Agresti, Michał Siemaszko, Joshua Morris, Borivoje Dakić, Riccardo Albiero, Andrea Crespi, Francesco Ceccarelli, Roberto Osellame, Magdalena Stobińska, Philip Walther · 2025
Read Experimental neuromorphic computing based on quantum memristor by Mirela Selimović, Iris Agresti, Michał Siemaszko, Joshua Morris, Borivoje Dakić, Riccardo Albiero, Andrea Crespi, Francesco Ceccarelli, Roberto Osellame, Magdalena Stobińska, Philip Walther (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-18694.txt.
Introduction
Mirela Selimović and ten co-authors build a brain-like computer from a quantum memristor in a 2025 paper. Neuromorphic computing copies synapses to learn with less energy. Quantum systems evolve in a linear way, which limits their nonlinearity. Nonlinear behavior is key to strong neural network work.
Their device is a photonic quantum memristor with a feedback loop. The loop boosts nonlinearity without entangling gates. That gain lifts the scores of the tested algorithms. Benchmarks cover a nonlinear function and three time series tasks. The memristive part proves vital in each benchmark. The team offers it as a block for larger quantum networks.
Catalog record
| CREATOR | Mirela Selimović, Iris Agresti, Michał Siemaszko, Joshua Morris, Borivoje Dakić, Riccardo Albiero, Andrea Crespi, Francesco Ceccarelli, Roberto Osellame, Magdalena Stobińska, Philip Walther |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 26 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-18694
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Experimental neuromorphic computing based on quantum memristor Mirela Selimović, Iris Agresti, Michał Siemaszko, Joshua Morris, Borivoje Dakić, Riccardo Albiero, Andrea Crespi, Francesco Ceccarelli, Roberto Osellame, Magdalena Stobińska, Philip Walther Abstract Machine learning has recently developed novel approaches, mimicking the synapses of the human brain to achieve similarly efficient learning strategies. Such an approach retains the universality of standard methods, while attempting to circumvent the excessive requirements, which hinder their scalability. In this context, quantum (or quantum inspired) algorithms may bring further enhancement. However, high-performing neural networks invariably display non-linear behaviours, which is diametrically opposed to the linear evolution of closed quantum systems. We propose a solution to this issue and report the first neuromorphic architecture based on a photonic quantum memristor. In detail, we show how the memristive feedback loop enhances the non-linearity and hence the performance of the algorithm. We benchmark our model on two tasks, i.e. non-linear function and time series prediction. In these cases, we highlight the essential role of the quantum memristive element and demonstrate the possibility of using it as a building block in more sophisticated networks. I Introduction In the last decades, machine learning has fundamentally changed the approach to solving a wide variety of problems, from everyday life to scientific scopes abdel-hamid_convolutional_2014 ; bakator_multimodal_2018 ; silver_mastering_2017 ; pierson_deep_2017 ; 755469 ; silver_mastering_2017 ; jumper_highly_2021 ; brown_language_2020 . In this context, artificial neural networks (ANN) have proven to be a pivotal tool, especially for tackling tasks where information needs to be extracted from high-dimensional data krizhevsky_imagenet_2017 . Nevertheless, a growing number of parameters must be trained, with correspondingly larger training data, as the tasks get more complex, resulting in the so-called memory bottleneck mutlu_processing_2019 . This sparked the development of more dedicated hardware, i.e. neuromorphic architectures, which mimic the human brain ham_neuromorphic_2021 . The main advantage of these models is their minimal learning requirements, which circumvents the aforementioned difficulty. Acting as the poster-child of neuromorphic components is the memrsitor, a passive circuital component that was postulated in 1971 chua_memristor-missing_1971 and demonstrated in 2008 strukov_missing_2008 . Its main characteristic is that it retains memory of its past states, in the form of a hysteresis that is similar to neural synapses jo_nanoscale_2010 ; du_reservoir_2017 ; ham_neuromorphic_2021 ; xiao_review_2023 ; zhong_dynamic_2021 . (a) (b) Figure 1 : Neuromorphic computing based on quantum memristor. (a) Non-linear activation functions in neural networks are crucial for learning input/output correlations. Indeed, linear functions limit the expressiveness of the model, not allowing for nuanced decisions. (b) The basic principle of a quantum memristor is a tunable Mach-Zehnder interferometer, where its internal phase ( \theta ) is varied, according to the output registered at one of its output modes, through a feedback loop. The investigated hybrid quantum/classical models are implemented through three parts: input encoding (quantum), non-linear reservoir (quantum) and one-layered ANN (classical). In our work, the encoding of classical variables in a one qubit state exploits the amplitude of the component parallel to |0\rangle , i.e. x\rightarrow\sqrt{x}|0\rangle+\sqrt{1-x}|1\rangle and x|0\rangle+\sqrt{1-x^{2}}|1\rangle . The reservoir consists in a unitary operation, composed by the product of different rotations. These rotations allow us to achieve a non-linear transformation of the input, due to the encoding that is non-linear in the density matrix of the quantum state. The loop enhances the achievable non-linearity, while implementing a short-term memory. The non-linear reservoir is then followed by a classical ANN, which is the only part of the model that is trained, producing a weighted sum of the reservoir outputs. The two addressed tasks are the prediction of a smooth non-linear function and of a random time series. (a) (b) Figure 2 : Experimental setup. (a) The experimental apparatus consists of a source of single photon pairs, which is pumped with a laser at 775 nm and generates degenerate photon pairs at 1550 nm. One photon is routed to the 3 input/output mode photonic integrated circuit (PIC) and the other photon is ancillary. Let us note that this circuit is implemented following the universal Reck architecture reck_experimental_1994 . At the outputs of the PIC, photon counts are registered by superconducting nanowire single photon detectors. After the full calibration of the chip, phases are set and modulated by applying a suitable voltage to the resistive elements, that act as thermal phase shifters. The unitary operation is updated, conditioned on the statistics of previous runs. The output distribution is then fed into a classical ANN, which is the only part of the machine learning model that is trained. (b) The PIC consists in 3 Mach-Zehnder interferometers and 5 tunable phases. The R_{i} indicate the resistive elements that serve as tunable thermal phase shifters, where R_{3} is updated conditioned on the output of the third mode, implementing a memristive behaviour. Let us note that the state obtained tracing out the third output still preserves coherence (see section II of the Supplemental Material). The input is encoded in one qubit, through a unitary transformation \textbf{U}_{enc} . The unitary transformations \textbf{U}_{1} and \textbf{U}_{2} are hyperparameters of our model. Then, \textbf{U}_{mem} amounts to the action of the memristor and is varied depending on previous outcomes. Also the feedback rule, according to which this happens, is a hyperparameter. A related but somewhat orthogonal paradigm shift in computer science was introduced by quantum computing. The interest in this field stems, first, from the possibility of tackling non-classical processing tasks directly related to the investigation of quantum effects and, second, from the promise of outperforming standard algorithms for particular problems grover_fast_nodate ; shor_polynomial-time_1997 ; montanaro_quantum_2016 ; harrow_quantum_2009 . However, for the latter, the only proven advantages were shown either for tasks which are out of reach of even near term state-of-the-art quantum computers or which have no known applications arute_quantum_2019 . Hence, a research line that has attracted a lot of interest arises from the combination of machine learning and quantum computation, i.e. Quantum Machine Learning (QML) dunjko_machine_2018 ; wittek2014quantum . This is fueled by the hope that a computational advantage can be found on practically relevant tasks even when using relatively small scale quantum information processing. Furthermore, photonic apparatuses exhibit vastly lower energy consumption, compared to standard electrical ones mcmahon_physics_2023 ; hamerly2019large , indicating the possibility of far more efficient computational platforms. However, in this context, the main difficulty lies in the deterministic implementation of non-linear operations (crucial for the learning process, see Fig. 1 a), since the evolution of closed quantum systems is intrinsically unitary (and therefore linear). Moreover, the introduction of interactions with the environment results in the reduction of the coherence of the overall system which, in turn, leads to a processing power that approaches the one of classical information processing. This is the reason for which employing Fock states for QML algorithms is challenging, although they can grant enhanced expressivity and they allow reaching higher-dimensional Hilbert spaces nerenberg2024photon ; gan2022fock ; yin_experimental_2024 with respect to (weakly)-coherent light inputs ghosh2019quantum ; henaff2024optical ; nokkala2021gaussian ; vandoorne_experimental_2014 . In this work, we propose a solution to this apparent conundrum and we implement the first instances of a neuromorphic computing protocol, i.e. reservoir computing (RC), on a photonic platform. We achieve this by exploiting single photon states manipulated through an integrated circuit mujal_opportunities_2021 . The model may be