Challenges and Opportunities in Quantum Machine Learning

• QML embeds machine learning and learning theory into quantum mechanics to study the limits of data analysis allowed by physics.Source: §I · p. 1
• Quantum computing may provide substantial speedups, but quantum advantage for data science remains theoretically uncertain.Source: §I · p. 1
• QML encompasses classical or quantum algorithms applied to classical or quantum data, including quantum-inspired classical methods and quantum neural networks.Source: §I · p. 1, Figure 1 · p. 2
• Potential applications include chemistry, materials science, sensing, metrology, classical data analysis, quantum error correction, and quantum algorithm design.Source: §I · p. 1, Figure 2 · p. 2
• Quantum hardware noise, including decoherence and shot noise, complicates QML training, while quantum linearity makes nonlinear model operations require careful design.Source: §I · pp. 1–2
• Demonstrating quantum advantage will require scalable, trainable models, suitable prediction-error analysis, and high-quality quantum hardware.Source: §I · p. 2
• Viewing quantum information, error correction, and foundations through data science may lead to new breakthroughs.Source: §I · p. 2

• Classical data includes images, texts, graphs, medical records, financial data, molecular properties, biological outcomes, and high-energy physics traces.Source: A. Data · p. 2
• Quantum data is encoded in qubits or higher-dimensional quantum systems and can represent states or unitaries associated with physical processes.Source: A. Data · pp. 2–3
• Classical bitstrings can be efficiently encoded on qubits, whereas general n-qubit states cannot generally be efficiently encoded in classical bits.Source: A. Data · pp. 2–3
• For classical inputs, an embedding maps x_j to quantum states |ψ(x_j)⟩ in Hilbert space; quantum inputs are already represented as states or state-preparing unitaries.Source: A. Data · p. 3
• The expected growth of quantum simulations and quantum computing may produce more quantum datasets, initially stored classically as efficient circuit descriptions.Source: A. Data · p. 3
• Improved quantum control, transduction, error correction, and quantum memories could support direct acquisition and storage of quantum information.Source: A. Data · p. 3

• Classical neural and tensor networks can analyze data from quantum experiments, while quantum models apply learning methods directly to quantum systems.Source: B. Models · p. 3
• QML uses supervised, unsupervised, and reinforcement learning paradigms for task-based, data-based, and reward-based learning, respectively.Source: B. Models · p. 3
• Supervised learning has attracted attention for potential quantum advantage, noise resilience, and generalization properties, making it a candidate for near-term applications.Source: B. Models · p. 3

1. Quantum neural networks

• Parameterized quantum circuits apply trainable unitary gates to quantum data states and provide the basic ingredient of QNN models.Source: §1 · p. 3
• QNN circuits propagate encoded information through layers of quantum gates before measuring some or all qubits for outputs.Source: Figure 4 · p. 4, §1 · p. 4
• QNNs can map different classes to distinguishable Hilbert-space regions and support supervised, unsupervised, and reinforcement learning.Source: §1 · p. 4
• Dissipative QNNs discard qubits and replace them across layers, standard QNNs preserve qubits, and convolutional QNNs measure qubits to reduce dimensionality.Source: §1 · p. 4, Figure 4 · p. 4
• Hybrid models combine coherent QNN processing with deep classical networks to accommodate limitations of near-term quantum computers.Source: §1 · p. 4

Quantum kernels

• Quantum kernel methods compute inner products using quantum computers, providing an alternative to quantum neural networks.Kernel methods learn linear functions in a potentially high-dimensional reproducing kernel Hilbert space.Source: §2 · p. 4

Inductive bias

• Inductive bias restricts the model search space toward selected functions and is a key design criterion for both quantum kernels and QNNs.Choosing a parameterization or embedding imposes assumptions about the functions the model can represent.Source: §3 · p. 5
• Quantum advantage may arise when a QML model has an inductive bias that is inefficient for classical models to simulate.Quantum kernels with this property have been constructed, although their trainability involves subtleties.Source: §3 · p. 5
• Quantum-native inductive biases are considered more promising for quantum data than for classical data.The paper connects this promise to models that can sample from and learn partially native quantum-mechanical structures.Source: §3 · p. 5

Training and generalization

• QML training commonly minimizes a loss function over parameters, but shot noise, hardware noise, and quantum landscape features complicate optimization.Quantum-aware methods include shot-frugal optimizers and quantum natural gradient.Source: C. Training and Generalization · p. 5
• Generalization error measures performance on unseen data and is small when training error is small and model complexity is moderate.The paper notes that quantum-kernel complexity is relatively well understood.Source: C. Training and Generalization · p. 5

III. CHALLENGES IN QML

• QML research aims to identify technical challenges quickly and focus effort on addressing them to avoid periods of stagnation.Figure 5 organizes model components alongside associated challenges.Source: §III · p. 5
• QML models require datasets, classical-data encoding schemes, parameterized models, loss functions, and classical optimizers.The figure presents these ingredients as interconnected components of model construction.Source: Figure 5 · p. 6

Embedding schemes and quantum datasets

• High-quality standardized datasets are expected to be important for advancing QML, as they have been for classical machine learning.The passage presents this as a conjecture rather than an established result.Source: A. Embedding schemes and quantum… · p. 6
• Most QML architectures are benchmarked on accessible classical datasets, but encoding classical information into quantum states remains unclear.Examples include MNIST, Dogs vs Cats, and Iris.Source: A. Embedding schemes and quantum… · p. 6
• Useful embeddings must combine classical hardness of simulating output-state inner products with distinguishable states for classification.An embedding satisfying one property does not necessarily satisfy the other.Source: A. Embedding schemes and quantum… · p. 6
• Encoding schemes with dataset-informed inductive bias are an active research area.The review identifies such bias as a desirable property for embedding schemes.Source: A. Embedding schemes and quantum… · p. 6
• Quantum advantage appears less straightforward with classical data, while quantum-data models have a more promising route but lack sufficient standardized datasets.The paper calls for datasets containing easily preparable quantum states for benchmarking.Source: A. Embedding schemes and quantum… · p. 6

B. Quantum landscapes

• Training QML parameters typically minimizes a loss function over a non-convex landscape in search of a global minimum.Quantum landscape theory studies these properties and how to engineer them.Source: B. Quantum landscapes · p. 6
• Many local minima can make overall quantum optimization NP-hard, as in the classical case.Variable-structure QNNs can adapt the model prior and turn some local minima into saddle points.Source: §1 · p. 6
• Overparametrization can produce a computational phase transition in which spurious local minima disappear beyond a critical parameter count.This phenomenon has been observed in QML.Source: §1 · p. 6
• In a barren plateau, the loss landscape becomes exponentially flat with problem size and the global-minimum valley shrinks into a narrow gorge.Navigating such a landscape requires exponential resources such as numbers of shots.Source: §2 · p. 6
• Barren plateaus can increase algorithmic complexity and even destroy quantum speedup by imposing exponential navigation resources.This conflicts with the goal of avoiding classical exponential complexity.Source: §2 · pp. 6–7
• Deep hardware-efficient QNNs develop barren plateaus partly because their highly expressive architectures use relatively uninformed priors.The phenomenon is linked to insufficient inductive bias.Source: §3 · p. 7
• Clever initialization, pre-training, parameter correlation, and sharper architectural priors are proposed strategies for mitigating barren plateaus.These methods constrain the search over over-expressive parameterizations.Source: §3 · p. 7

4. Barren plateaus from global observables

• Global-observable loss functions can cause barren plateaus even in shallow circuits with sharp priors.The issue arises when observables measure all qubits.Source: §4 · p. 7
• Local observables avoid the global-observable barren-plateau issue by comparing quantum states at the single-qubit level.The distinction reflects the difficulty of comparing objects in exponentially large Hilbert spaces.Source: §4 · p. 7
• Excessive entanglement between visible output qubits and many hidden qubits can also produce barren plateaus.Information becomes distributed across non-local correlations, leaving visible-qubit reduced states near maximally mixed.Source: §5 · p. 7
• Taming entanglement across the QNN is proposed as a way to solve this entanglement-induced barren plateau.This mechanism differs from barren plateaus caused by inadequate inductive bias.Source: §5 · p. 7

Architecture principles

• Sharp priors are a central design goal because QNNs support supervised, unsupervised, and reinforcement learning.The paper identifies architecture design as crucial across these learning settings.Source: C. QNN architecture design · p. 7
• QCNNs restrict architectures to hierarchical and translationally invariant structures, reducing expressivity and parameter-space dimensionality.These sharper priors have been shown to avoid barren-plateau issues altogether.Source: C. QNN architecture design · p. 7

Symmetry-informed models

• Embedding dataset and problem knowledge, including graph-permutation symmetries, can provide useful inductive bias for QML models.Quantum graph neural networks and graph convolutional networks are examples of this strategy.Source: C. QNN architecture design · p. 7
• Classical geometric deep learning offers group-theoretic concepts of invariance and equivariance as inspiration for quantum architectures.This direction seeks to move beyond inductive biases based only on temporal or translational invariance.Source: C. QNN architecture design · p. 7

Quantum geometric deep learning

• Quantum geometric deep learning could prescribe architectures whose transformation-space priors encode symmetries and principles of quantum physical systems.The intended benefits are improved trainability and generalization, with related approaches recently explored.Source: C. QNN architecture design · pp. 7–8

Noise effects

• Hardware noise and measurement shot noise complicate QML training, while quantum linearity makes nonlinear operations require careful model design.The relevant hardware effects include decoherence and statistical noise from measurements.Source: D. Effect of quantum noise · p. 8
• Noise propagates through quantum circuits, disproportionately affecting deeper and longer-running circuits and potentially inducing barren plateaus in QNNs.It can also affect data preparation and quantum-kernel computation.Source: D. Effect of quantum noise · p. 8

Mitigation strategies

• Mitigating QML noise may require lower hardware error rates, partial quantum error correction, or shallow QNNs such as QCNNs.Error mitigation can improve performance but may not resolve noise-induced trainability problems.Source: D. Effect of quantum noise · p. 8

Potential advantages

• The first QML advantages are expected from hidden-parameter extraction in quantum sensing and quantum state classification or regression.Non-local measurements can extract hidden parameters using fewer samples, and QML can search over parameterized measurement hypotheses.Source: A. Potential for Quantum advantage · p. 8
• QML may also advantage quantum error-correcting-code discovery and generative modeling of ground states, equilibrium states, or quantum dynamics.Generative advantage requires distributions that cannot be sampled classically and improved prediction or generalization relative to classical ML.Source: A. Potential for Quantum advantage · pp. 8–9

Quantum-origin data

• Quantum-mechanical data from chemistry, materials science, biology, and physics is considered more likely to yield exponential QML advantage.Potential gains may concern either sample complexity or time complexity.Source: B. What will quantum advantage look… · p. 9
• An exponential sample-complexity advantage has been rigorously shown when quantum sensors, memories, and computers retrieve, store, and process experimental quantum information.The passage reports that this advantage has also been demonstrated on the Sycamore processor.Source: B. What will quantum advantage look… · p. 9

Time complexity

• Exponential time-complexity advantage is more subtle because data availability can make classical ML algorithms computationally more powerful.The paper notes no rigorous proofs yet for an exponential advantage in practical physical systems.Source: B. What will quantum advantage look… · p. 9

Classical-origin data

• For purely classical data, no exponential QML advantage is known, although polynomial and rigorously proven quadratic advantages remain possible.Potential long-term impact is constrained by the overhead of quantum error correction in fault-tolerant schemes.Source: B. What will quantum advantage look… · p. 9

Post-NISQ eras

• The partial-error-corrected era combines a small number of logical qubits with non-error-corrected physical qubits, enabling a trade-off between the two.The fault-tolerant era begins when quantum hardware has many error-corrected qubits.Source: C. Transition to the fault-tolerant… · p. 9

Partial error correction

• High-fidelity circuits in the partial-error-corrected era could improve performance, mitigate noise-induced barren plateaus, and reduce noise-induced classification errors.These effects are presented as reasons QML may be useful before fully fault-tolerant hardware.Source: C. Transition to the fault-tolerant… · p. 9

Fault tolerance

• During the fault-tolerant era, quantum simulation and quantum memories could prepare and store quantum data for QML models to learn directly from it.The paper describes QML as a natural model for learning, inference, and prediction from such data.Source: C. Transition to the fault-tolerant… · p. 9

Beyond fault tolerance

• Further-term QML could receive quantum data captured from nature through quantum analog-digital interconversion and distributed over quantum networks.This scenario is compared with contemporary classical ML pipelines that relay sensor data to centralized processing.Source: C. Transition to the fault-tolerant… · p. 9