Session Y50: Quantum machine learning - Near-Term Applications

Practical hybrid digital-analog quantum learning on Rydberg atom arrays

We propose hybrid digital-analog learning algorithms on Rydberg atom arrays, combining the

potentially practical utility of quantum learning and near-term realizability with the rapidly scaling

architectures of neutral atoms. Our construction requires only single-qubit operations in the digital

setting and global driving according to the Rydberg Hamiltonian in the analog setting; this “digital-

analog” scheme has already been achieved in recent experiments and will become widely available in

the coming years. We perform a comprehensive numerical study of our algorithms on both classical

and quantum data, given respectively by a modified version of handwritten digit classification and

unsupervised quantum phase boundary learning. We show in the two representative problems that

digital-analog learning is not only more feasible in the near term, but also requires shorter circuit

depths and is more robust to realistic error models as compared to digital learning schemes. These

advantages can be explained by simple physical principles. Our results suggest that digital-analog

learning opens a promising path towards practical algorithm implementations in the near term, and

raises new questions about the power and limits of digital-analog circuitry for quantum learning.   

Quantum-enhanced machine learning using phosphorus-doped silicon quantum dots

Recently, quantum machine learning has garnered intense interest due to the promise for improvements over classical ML and its applicability on near-term hardware. In particular, recent theoretical work has shown that quantum kernel functions admit a provable quantum advantage over classical kernels. Quantum kernels are often used as part of variational hybrid quantum-classical machine learning algorithms that are within the reach of NISQ computers. Typically, the optimisation of ansatz circuits used in this variational technique is costly, requiring significant time and often running into barriers such as the barren plateau problem. Randomisation can be used as an alternative to optimisation in this context. By creating an enhanced feature space using randomness, comparable results to kernel functions can be achieved in a shorter time. In this talk, I will show how we can leverage atomically precise manufacturing of phosphorus dopants in silicon to realise a quantum system for generating features for use in a random quantum feature algorithm. We experimentally demonstrate the algorithm on various datasets and find the performance is competitive with equivalent classical methods. The results indicate that this is a promising approach for achieving near-term quantum advantage.   

Utilizing the non-interacting bionic particle sampling to solve image classification tasks

The sampling of non-interacting boson particles has been shown to be a hard classical problem to solve. It is known as the #P class in computational complexity theory and tells us that simple systems may be used as a computational resource. It however does not tell us how to utilize them. Doing so is nontrivial. Using non-interacting boson particles, we introduce a new approach based on the quantum neural network model. The presentation discusses the design of the encoder, reservoir, and measurement process. The best performance with our models achieves a 96.6% accuracy rate for testing of hand-written digit images (MNIST) with 4 photons and 16 waveguides, which is within the reach of the current technology.   

Quantum-enhanced physical-layer data learning with a variational sensor network

The emergence of quantum sensor networks has opened up the opportunity to enhance complex sensing tasks that routinely arise in practice. However, it has also brought tremendous challenges in designing and analyzing optimal quantum sensing protocols. Specifically, when it comes to hypothesis testing between physical-layer data modeled as quantum channels, the analytical approach is only sufficient to handle the simple linearly separable case. In this case, the error probability is reduced through Gaussian entanglement and measurement, as theoretically predicted in [Phys. Rev. X 9, 041023 (2019)] and experimentally verified in [Phys. Rev. X 11, 021047 (2021)]. This leaves an open question in the context of general nonlinear physical-layer data classification tasks.  In this work, we develop supervised learning assisted by an entangled sensor network (SLAEN) for nonlinear classification. Empowered by universal quantum control readily available in cavity-QED experiments, we take a variational approach to train SLAEN to achieve the optimal advantage. In linearly separable tasks, we identify a threshold phenomenon in the classification error, where the error abruptly decreases to near-zero at a specific probe energy threshold. This results in a substantial advantage over classical or the previous Gaussian SLEAN. Despite the non-Gaussian nature of the problem, we provide analytical analyses to determine the threshold and residual error. In the case of nonlinear data, we also identify a significant advantage over classical or Gaussian strategies. Our findings have implications in the fields of radio-frequency photonic sensors and microwave dark matter haloscopes.

* NSF, DARPA, ONR.   

Quadri-partite Quantum-Assisted VAE as a calorimeter surrogate

Simulations of collision events at experiments like ATLAS and CMS have played a pivotal role in shaping the design of future experiments and analyzing ongoing ones. However, the quest for accuracy in describing Large Hadron Collider (LHC) collisions comes at an imposing computational cost, with projections estimating the need for millions of CPU-years annually during the High Luminosity LHC (HL-LHC) run [1]. Simulating a single LHC event with Geant4 currently devours around 1000 CPU seconds, with calorimeter simulations imposing substantial computational demands [2]. To address this challenge, we propose a Quantum-Assisted deep generative model. Our model marries a variational autoencoder (VAE) on the exterior with a Restricted Boltzmann Machine (RBM) in the latent space, delivering enhanced expressiveness compared to conventional VAEs. The RBM nodes and connections are meticulously engineered to enable the use of qubits and couplers on D-Wave's Pegasus Quantum Annealer for sampling and training. We also provide preliminary insights into the requisite infrastructure for large-scale deployment.

[1] A. Collaboration, ATLAS software and computing HL-LHC roadmap, Tech. Rep. (Technical report, CERN,Geneva. http://cds. cern. ch/record/2802918, 2022).

[2] D. Rousseau, Experimental particle physics and artificial intelligence, in Artificial Intelligence for Science: A Deep Learning Revolution (World Scientific, 2023) pp.447–464   

Quantum Hardware-Enabled Molecular Dynamics via Transfer Learning

The ability to perform ab initio molecular dynamics simulations using potential energy surfaces provided by quantum computers would open the door to virtually exact dynamics for a variety of chemical and biochemical systems, with impacts on catalysis and biophysics. Nonetheless, performing molecular dynamics on surfaces produced by quantum hardware has been hampered by the noisy energies typically produced by quantum computers and challenges associated with computing gradients and scaling to large systems interest. A recent set of advances in machine learning, known as transfer learning, provides a new path forward for molecular dynamics simulations on quantum hardware. Transfer learning offers a workaround, where one first trains models on larger, less accurate classical datasets and then refines them on smaller, more accurate quantum datasets. We explore this approach by training machine learning models to predict a molecule's potential energy based on its geometric structure using Behler-Parrinello neural networks. When successfully trained, the model enables energy gradient predictions necessary for dynamic simulations. To reduce the quantum resources needed, the model is initially trained with data derived from classical density functional theory and subsequently refined with a smaller dataset obtained from a variational quantum eigensolver optimization of the unitary coupled cluster ansatz. We show that this approach significantly reduces the size of the needed quantum training dataset while capturing the high accuracies needed within quantum chemistry simulations. The success of this two-step training method opens more opportunities to apply machine learning models on quantum data, a significant stride towards efficient quantum-classical hybrid computational models.   

Machine Learning with Near-Term Quantum Computers

Recent studies on quantum computing simulators and hardware indicate that parametrized quantum circuits used for learning on classical data can achieve results similar to that of classical machine learning models while using significantly fewer parameters. We will present results from these studies in a variety of application areas ranging from image recognition to modeling financial data on up to 20 qubits. We will then show that the origin of this quantum advantage can be related to the hardness of modeling correlations present on small scales or in the tails of multivariate data distributions. In particular, we show that quantum computers can model arbitrary multivariate distributions with a number of parameters that scale linearly in the number of variables (features), whereas models used in classical machine learning will typically use an exponential number of parameters. We show that these arguments holds for both discriminative and generative learning, and leads to quantum models that are suitable for near-term quantum computers. We will also propose a metric that can a priori identify suitability of a particular dataset for either classical or quantum machine learning models.   

Quantum Reservoir Computing with Neutral Atom Arrays

With the development of quantum computers, quantum machine learning has recently attracted much attention. While it has been considered a promising application for near-term quantum computers, current quantum machine learning methods require large quantum resources and suffer from gradient vanishing issues. To alleviate this, we propose a general-purpose quantum reservoir computing algorithm for neutral atom quantum simulators that is resource-frugal, noise-resilient, and scalable. We implement our proposal on QuEra's field-programmable qubit array, Aquila, and observe state-of-the-art performance on several practical machine-learning tasks.   

A Practical Quantum Reservoir Computing Platform for Quantum Data Processing

Quantum Reservoir Computing (QRC) has the potential to combine the low-latency and energy-efficient machine learning of classical RC with the computing power of complex and high-dimensional quantum dynamics. This is a particularly compelling framework for current quantum hardware, as QRC requires neither complex controls or highly-calibrated operations. Here we present a novel superconducting circuit architecture for QRC, and describe its application to meaningful tasks involving the processing of time-dependent and quantum data, in both simulation and experiment. The device is comprised of an array of transmons coupled to a flux-tunable bus resonator which both mediates all-to-all coupling and provides a robust readout channel. By varying the probe strength and flux bias, the reservoir can be tuned to various processing regimes, enabling a trade-off between nonlinearity and memory capacity through a unified QRC framework. We present a QRC implementation of quantum state tomography, which is performed with near-optimal measurement resources and without the need to calibrate any gate sequences. Our QRC platform and framework poses a practical solution to a wide range of quantum information processing tasks.   

Practical and versatile reservoir-filter for optimizing multi-state qubit readout

Qubit state readout has relied on matched filters to extract information from measurement data, in spite of their applicability only for white noise and binary classification. We demonstrate a reservoir-computing inspired learning scheme [1] for optimal temporal processing of quantum measurement data dominated by noise of a quantum-mechanical origin, such as quantum jumps or added noise of quantum amplifiers, for classification of an arbitrary number of states. Through demonstrations on real qubits, we show that reservoir classification can outperform standard approaches in complex readout regimes at high readout powers with multi-level transitions. Via a heuristic interpretation of reservoir learning as optimal filtering, we show that reservoir-learned filters account for correlations in data, including those due to quantum noise. For white noise, our approach provides a generalization of binary matched filtering to an arbitrary number of states. More importantly, across both experiments and theoretical simulations with quantum noise sources such as amplifier added noise and quantum jumps, we show the reservoir can uncover optimal filters that outperform matched filtering. The reservoir framework requires only linear weights and is thus ideal for real-time processing via FPGAs.  

[1] Git repo: https://github.com/skhanCC/multistateRC-py   

Reservoir Computing with Feedback for Quantum State Identification

Identifying quantum states through measurements is a complex challenge encountered in many quantum applications. Given ample time and multiple measurements, quantum tomography is frequently used to pinpoint the state. However, real-world qubit readouts are often constrained by the information acquired from specific read-out setups. While neural networks have been used to address quantum state identification, their creation and operation come at a significant computational expense. Reservoir computing offers a potential solution by substituting a naturally occurring nonlinear dynamical system for the neural network. This switch reduces training costs while preserving functional versatility.

 

Echo state networks (ESNs) are a subset of reservoir computers. In ESNs, the internal reservoir state undergoes a linear transformation at each timestep, which is both fixed and randomly chosen. Subsequently, this state is subject to a nonlinear transformation. The reservoir dynamics then guide the system towards a sequence of states dictated by the input signal. This sequence can then be tailored to a linear function, encapsulating the characteristics of the desired target system. A potential drawback is that the fixed reservoir states might not offer the complexity needed for specific problems. While directly altering (training) the ESN would reintroduce the computational burden, an indirect modification can be achieved by redirecting some output as input. This feedback loop can potentially influence the internal reservoir state, yielding enhanced ESN outputs suitable for a broader array of challenges.

 

We suggest deploying an ESN, with output feedback, to interpret quantum state readouts from partial or incomplete measurements. Given the intrinsic quantum nature of the problem, our proposal leans towards a quantum solution: a quantum ESN. This approach seems optimal, as replicating quantum states on classical computing systems is notoriously computationally intensive. Thus, a quantum-based ESN should be more efficient for quantum state readouts. By feeding the measurements from a disordered quantum state, we aim to calibrate our feedback-augmented ESN to approximate the original state closely.   

Hierarchy of the echo state property in quantum reservoir computing

The echo state property (ESP) represents a fundamental concept in the reservoir computing framework that ensures stable output-only training of reservoir networks. However, the conventional definition of ESP does not aptly describe possibly non-stationary systems, where statistical properties evolve. 

To address this issue, we introduce two new categories of ESP: non-stationary ESP designed for possibly non-stationary systems, and subspace/subset ESP designed for systems whose subsystems possess ESP. In the quantum reservoir computing (QRC) framework under typical Hamiltonian dynamics and input encoding methods, we numerically demonstrate the correspondence between non-stationary ESP and the performance of nonlinear autoregressive moving average (NARMA) and linear/nonlinear memory tasks. Our findings reveal that the tendencies observed in the root normalized mean squared error (RNMSE) of the NARMA task closely aligned with the distribution of the non-stationary ESP indicator. Furthermore, metrics such as total memory capacity (MC) and total information processing capacity (IPC), derived from linear and nonlinear memory tasks, exhibit behaviors consistent with non-stationary ESP with partial exceptions. It suggests that the non-stationary ESP explains at least a large part of the system's information processing capability. Our proposed categories of ESP provide a new understanding of the practical design of QRC, and other RC systems that may exhibit some type of non-stationarity.