Bulletin of the American Physical Society
APS March Meeting 2022
Volume 67, Number 3
Monday–Friday, March 14–18, 2022; Chicago
Session W37: Quantum Machine Learning IIIFocus Recordings Available
|
Hide Abstracts |
Sponsoring Units: DQI GDS Chair: Pranav Rao, UIUC Room: McCormick Place W-194B |
Thursday, March 17, 2022 3:00PM - 3:12PM |
W37.00001: Entangled Datasets for Quantum Machine Learning Louis Schatzki, Andrew T Arrasmith, Patrick J Coles, Marco Cerezo High-quality, large-scale datasets played a crucial role in the development and success of classical machine learning. Quantum Machine Learning (QML) is a new field that aims to use quantum computers for data analysis, with the hope of obtaining a quantum advantage. While most proposed QML architectures are benchmarked via classical datasets, there are still doubts if QML on classical datasets can achieve an advantage. In this work, we argue that one should instead employ quantum datasets composed of quantum states. For this purpose, we introduce the NTangled dataset composed of states with different amounts and types of multipartite entanglement for up to 12 qubits. We first show how a quantum neural network (QNN) can be trained to generate the states in the dataset. Then, we use the NTangled dataset to benchmark QML models for supervised learning tasks. We also consider an alternative scalable entanglement-based dataset composed of states prepared by circuits of different depths. With QNNs, we demonstrate that high classification accuracies can be achieved for learning the depth of an ansatz. As a byproduct of our results, we introduce a novel method for generating multipartite entangled states, providing a use-case of quantum neural networks for quantum entanglement theory. |
Thursday, March 17, 2022 3:12PM - 3:24PM |
W37.00002: Quantum Receiver Enhanced by Adaptive Learning Chaohan Cui, William Horrocks, Saikat Guha, Nasser Peyghambarian, Quntao Zhuang, Zheshen Zhang Quantum receivers are essential components for quantum information processing (QIP), aimed at capturing information embedded in quantum states far more efficiently than any conventional classical receiver allows. To date, only a handful of quantum-receiver structures have been laid out to approach the ultimate performance bounds in their respective QIP tasks, and their advantage in realistic operational environments is hindered by noise, turbulence, and other imperfections. Moreover, a traditional analytic approach is incapable of customizing quantum-receiver structures due to the vast parameter space of the optimization. |
Thursday, March 17, 2022 3:24PM - 3:36PM |
W37.00003: QML Suite: Quantum Machine Learning for Large-Scale Applications Brian J Dellabetta In this talk, we introduce QML Suite: Zapata Computing's library to support large-scale quantum machine learning tasks in application domains ranging from quantum generative modeling, combinatorial optimization, and the optimization and visualization of parametrized quantum circuits. We demonstrate composability, highlighting how components in hybrid architectures can be swapped for other implementations or to compare against classical benchmarks. We present three applications that leverage quantum-inspired or hybrid quantum-classical architectures, highlighting the potential regimes in which an advantage is achieved by means of our techniques and the flexibility of our platform. We present results obtained with QML Suite to assist the solution of industrial-scale combinatorial optimization problems with tensor-network-based generative models, the design of quantum-assisted generative adversarial networks and a strategy to mitigate barren plateaus with a flexible meta-learning initialization of parameters in variational quantum algorithms. |
Thursday, March 17, 2022 3:36PM - 3:48PM |
W37.00004: Quantum Annealing Systems as Reservoirs I: Design and Performance Gerasimos M Angelatos, Hakan E Tureci, Fangjun Hu, Saeed A Khan Modern quantum processors offer an unprecedented number of quantum degrees of freedom which can be controlled and read-out. Although NISQ constraints limit the utility of such devices for conventional quantum machine learning applications, they can readily serve as the physical platform of a reservoir computer or extreme learning machine. Here we propose a general algorithm to operate an existing Quantum Annealing System as a Reservoir (QASAR), and discuss its application to static pattern recognition and dynamic signal processing tasks, including signal classification and channel equalization. We find robust performance across a variety of tasks, with state-of-the-art resource efficiency in terms of qubits, circuit runs, and training overhead. Unlike conventional quantum machine learning approaches, the dissipation in these fundamentally open systems is an essential ingredient to successful signal processing. We demonstrate how to readily increase this device's feature size and consequent performance through spatial multiplexing and the measurement of higher moments, at the cost of only a modest increase in circuit runs. |
Thursday, March 17, 2022 3:48PM - 4:00PM |
W37.00005: Quantum Annealing Systems as Reservoirs II: Quantum Channels and Computational Capacity Fangjun Hu, Gerasimos M Angelatos, Saeed A Khan, Hakan E Tureci Reservoir computing exploits the nonlinear dynamics and high dimensionality of physical systems to continually process time-dependent signals. Recent efforts have utilized conventional quantum processors as powerful and exotic dynamical maps for quantum reservoir computing (QRC). However, one needs to continuously extract information from this computational system, and the critical role of measurement in this nascent field is relatively unexplored. Here we present a fundamental analysis of general quantum circuit reservoirs under repeated measurements, a particular case of which was considered in Part 1. We find that the presence of a quantum channel is essential to imbue the reservoir with fading memory and avoid thermalizing due to repeated measurements. The simplest description of a quantum channel in this framework is the deterministic reset of a subset of the qubits after measurement through classical control operation, which is readily implementable in quantum processors. We evaluate the fundamental information processing and memory capacity of our proposed QASAR computing framework, exemplifying the efficacy of this dephasing-resistant approach and additionally demonstrating its robust processing ability in the presence of noise and finite sampling. |
Thursday, March 17, 2022 4:00PM - 4:36PM |
W37.00006: Smart Superresolving Quantum Cameras Invited Speaker: Omar Magana-Loaiza The manifestation of the wave nature of light through diffraction imposes limits on the resolution of optical imaging. For over a century, the Abbe-Rayleigh criterion has been utilized to assess the spatial resolution limits of optical instruments. Recently, there has been an enormous impetus in overcoming the Abbe-Rayleigh resolution limit by projecting target light beams onto spatial modes. These conventional schemes for superresolution rely on a series of spatial projective measurements to pick up phase information that is used to boost the spatial resolution of optical systems. Unfortunately, these schemes require a priori information regarding the coherence properties of the, in principle, "unknown" light beams. Furthermore, they require stringent alignment and centering conditions that cannot be achieved in realistic scenarios. Here, I will describe our recent smart quantum cameras for superresolving imaging. These cameras exploit the self-learning features of artificial intelligence to identify the statistical fluctuations of unknown mixtures of light sources at each pixel. This is achieved through a universal quantum model that enables the design of artificial neural networks for the identification of quantum photon fluctuations. Our cameras overcome the inherent limitations of existing superresolution schemes based on spatial mode projection. Thus, our work provides a new perspective in the field of imaging with important implications for microscopy, remote sensing, and astronomy. |
Thursday, March 17, 2022 4:36PM - 4:48PM |
W37.00007: Reservoir Computing on Current Quantum Processors Marti Vives, Gerasimos M Angelatos, Fangjun Hu, Hakan E Tureci The Reservoir Computing (RC) paradigm enables the encoding of computation into the natural time evolution of any physical system that is sufficiently complex. Quantum circuits operated as Reservoir Computers (QRC) have recently been shown to have the potential to realize universal approximators for arbitrary functions and causal filters on input signals. In this work we develop a practical gate-based realization of a reservoir computer accounting for the role of dissipation and finite sampling via projective measurements. We focus on processing large classical datasets with current small-scale quantum hardware, enabled through an algorithm to vary input encodings and gate sets associated with reservoir evolution, as well as via the measurement of higher order moments. We apply this reservoir computing approach to nontrivial machine learning tasks such as handwritten digit recognition and dynamical signal processing, both through simulation and on IBMQ processors. By benchmarking against linear baselines, we demonstrate that the expressive power of the nonlinearity induced by the quantum hardware enables compelling performance using limited quantum resources, and without the need for error mitigation techniques. |
Thursday, March 17, 2022 4:48PM - 5:00PM |
W37.00008: Computational processing capacity of quantum reservoirs across the classical-to-quantum transition Saeed A Khan, Fangjun Hu, Gerasimos Angelatos, Hakan E Tureci Quantum reservoir computing is a resource-efficient machine-learning paradigm ideally suited to near-term quantum implementations, due to its relaxed requirements on hardware control and training volume. An important open question is the possible computational advantage of a quantum reservoir in comparison to its classical counterpart. We address this question directly by introducing a framework for quantum reservoir computing [1] that enables the same physical reservoir to be operated in quantum or classical regimes. Our approach is built upon a quantum-mechanical description of the complete measurement chain including a nonlinear multimode physical reservoir, taking into account realistic input schemes and quantum measurement overhead, realizable in the superconducting circuit architecture. We define a metric for information processing capacity which can be used to compare a physically-motivated classical limit of reservoir operation at high-excitation powers, to a well-defined quantum limit, thus isolating the role played by quantum dynamics. We also analyze the impact of quantum correlations and entanglement as resources for reservoir computing using multimode quantum reservoirs. |
Thursday, March 17, 2022 5:00PM - 5:12PM |
W37.00009: Quantum reservoir neural network implementation on a Josephson parametric converter Danijela Markovic, Julien Dudas, Julie Grollier Neuromorphic computing implements neural networks in hardware to make their training more time and energy efficient. However, addressing state-of-the-art machine learning tasks requires coupling large numbers of neurons, which is challenging with physical nano-devices. It has been proposed to solve this problem using quantum hardware and encoding neurons in the basis states, whose number is exponential in the number of coupled qubits. In this work, in order to obtain an even larger number of basis states, we use quantum oscillators instead of qubits. To go towards an experimental realization, we simulate a reservoir neural network implemented on a Josephson parametric converter. This circuit couples two superconducting oscillators through a three-wave-mixing interaction, implemented using a ring of four Josephson junctions. We encode the input data in the resonant oscillators' drives and numerically integrate quantum master equation to find the occupation probabilities of a subset of basis states that represent neural network outputs. We show that this system of two coupled quantum oscillators can solve a sine and square waveform classification task that otherwise requires 25 classical oscillators. Furthermore, we train this network to perform chaotic Mackey-Glass series prediction and show that with typical experimental parameters for a Josephson parametric converter we can obtain memory and performance comparable to other physical neural networks.These results show that a simple and well known quantum circuit can realize non-trivial machine learning tasks when its dynamics is exploited. Neuromorphic computing thus promises to leverage the full computing capabilities of even small quantum systems. These simulations will guide experimental realization of a reservoir neural network on the Josephson parametric converter. |
Thursday, March 17, 2022 5:12PM - 5:24PM |
W37.00010: Real-time adaptive Bayesian Estimation demonstrated on a non-single-shot readout sensor Inbar Zohar, Amit Finkler, Yoav Romach, Nir Halay, Cristian Bonato, Niv Drucker, Yonatan Cohen, Muhammad Junaid Arshad Quantum phase-estimation algorithms enhance the sensitivity and dynamic range of sensing protocols. Adaptive Bayesian Estimation (ABE) is one example of a method that has been proven to enhance single-shot readout (SSR) sensors. Nevertheless, many sensors do not always have an SSR capability and require averaged readout, where a threshold can be used to determine a binary result according to the number of positive measurements. This outcome is used in the likelihood function of the subsequent Bayesian update. Another approach considers the number of positive results from R measurements and uses a binomial distribution, that yields a different value for each outcome. The latter approach increases the amount of information provided. |
Thursday, March 17, 2022 5:24PM - 5:36PM |
W37.00011: Quantum machine learning with linear optics and coherent states Beng Yee Gan, Daniel Leykam, Dimitris G Angelakis Artificial optical neural networks based on coherent light propagation can potentially provide a higher information processing speed and lower power consumption than conventional electronic architectures [1]. Yet, their performance is limited by the difficulty of realizing nonlinear activation functions. In contrast, quantum machine learning approaches circumvent the need for nonlinearities by embedding data into high dimensional Hilbert spaces using quantum feature maps [2,3]. Here we show how the quantum feature map approach allows one to carry out photonic machine learning using coherent states, linear optical interferometers, and single photon or photon number-resolving detectors [4]. Our work sheds some light on the possibility of performing non-trivial quantum machine learning tasks using bosonic modes. |
Thursday, March 17, 2022 5:36PM - 5:48PM |
W37.00012: Quantum federated learning through blind quantum computing Weikang Li, Sirui Lu, Dong-Ling Deng Private distributed learning studies the problem of how multiple distributed entities collaboratively train a shared deep network with their private data unrevealed. With the security provided by the protocols of blind quantum computation, the cooperation between quantum physics and machine learning may lead to unparalleled prospects for solving private distributed learning tasks. Here, we introduce a quantum protocol for distributed learning that is able to utilize the computational power of the remote quantum servers while keeping the private data safe. For concreteness, we first introduce a protocol for private single-party delegated training of variational quantum classifiers based on blind quantum computing and then extend this protocol to multiparty private distributed learning incorporated with differential privacy. We carry out extensive numerical simulations with different real-life datasets and encoding strategies to benchmark the effectiveness of our protocol. We find that our protocol is robust to experimental imperfections and is secure under the gradient attack after the incorporation of differential privacy. Our results show the potential for handling computationally expensive distributed learning tasks with privacy guarantees, thus providing a valuable guide for exploring quantum advantages from the security perspective in the field of machine learning with real-life applications. |
Follow Us |
Engage
Become an APS Member |
My APS
Renew Membership |
Information for |
About APSThe American Physical Society (APS) is a non-profit membership organization working to advance the knowledge of physics. |
© 2024 American Physical Society
| All rights reserved | Terms of Use
| Contact Us
Headquarters
1 Physics Ellipse, College Park, MD 20740-3844
(301) 209-3200
Editorial Office
100 Motor Pkwy, Suite 110, Hauppauge, NY 11788
(631) 591-4000
Office of Public Affairs
529 14th St NW, Suite 1050, Washington, D.C. 20045-2001
(202) 662-8700