Quantum Machine Learning
Expert Perspectives on the Trajectory and Challenges of Quantum Machine Learning
Experts in quantum machine learning (QML) discuss the field's current state, emphasizing the need for critical evaluation beyond "speedups." Key areas of focus include the potential for QML algorithms to process quantum data directly, the importance of rigorous theoretical foundations, and the neces…
Quantum Computers Excel at Spectral Methods Fundamental to Machine Learning
Spectral methods, which manipulate the Fourier spectrum of ML models for learning and regularization, align naturally with quantum computing capabilities like the Quantum Fourier Transform. Representing generative models as quantum states enables efficient spectrum manipulation unavailable classical…
Quantum Fourier Transform Enables Exact Probabilistic Modeling of Permutations
Quantum computers exploit super-exponential speedup in the Quantum Fourier Transform (QFT) over the symmetric group to encode exact non-Abelian harmonic analysis models for permutation-structured data. These models, intractable classically, capture permutation correlations via group Fourier spectra,…
Quantum Hybrid SVMs Outperform Classical Methods in Wearable-Based Stress Detection for Older Adults
Quantum hybrid support vector machines frame stress detection as anomaly detection using smartwatch sensor data, with cortisol as ground truth and TSST protocol validation on 40 older adults. Kernel-based quantum preprocessing explores complex feature spaces efficiently with limited features. Experi…
Quantum Fourier Transform Inspires Symmetry-Leveraging Heuristics for Data Inference
The paper reinterprets the Hidden Subgroup Problem (HSP) as a machine learning task by replacing quantum oracles with classical training data. It leverages the Quantum Fourier Transform (QFT) from HSP solutions to identify invariant subspaces where hidden symmetries are exposed. An inference princip…
Quantum Non-Commutativity Enables Learning Survey Order Effects
Quantum models leverage non-commuting observables to capture order effects in data, such as varying human survey responses to question order swaps. A generative model of sequential learnable measurements adapts to tasks by reordering observables in a multi-task setup. Simulations on psychology-inspi…
Rethinking Quantum Machine Learning Research Paradigms
Current quantum machine learning research exhibits problematic patterns: superficial quantum adaptations of classical ML models, overemphasis on exponential speedups in unsuitable contexts, and a positivity bias in reporting benchmarks on small datasets. A new research agenda is needed to focus on t…
Quantum Models Replicate Benign Overfitting via Interpolating Fourier Features
Quantum machine learning models exhibit benign overfitting, generalizing well despite overfitting noisy training data, analogous to classical deep networks. The paper derives behavior in classical interpolating Fourier features models for noisy signal regression and maps it to quantum models, linkin…
Quantum Advantage Misguides Quantum Machine Learning Research
Classical ML algorithms excel in practice but resist theoretical analysis, while quantum computing lacks scalable benchmarks and relies heavily on theory for relevance assessment. Current tools fail to reliably evaluate quantum computers' practical utility for ML tasks. The field should shift from d…
Rethinking Quantum Advantage in Quantum Machine Learning Applications
The pursuit of "quantum advantage" as the primary driver for quantum machine learning (QML) development may be misapplied. Current quantum computers (NISQ devices) operate under different paradigms than theoretical fault-tolerant quantum computers, making traditional complexity theory an unsuitable …
Quantum Annealing Matches Classical Feature Selection for Stress Detection but Excels Under Data Scarcity
Quantum annealing selects optimal feature subsets from physiological signals (foot/hand EDA, ECG, respiration) for stress detection by embedding Pearson correlations into a binary quadratic model solved via D-Wave's clique sampler. It performs equivalently to classical methods under normal condition…
Quantum Annealing on D-Wave Outperforms Classical ML in Data-Limited, High-Dimensional Classification Tasks
Quantum annealing via D-Wave systems optimizes machine learning pipelines, particularly for classification in real-world applications constrained by limited training data and high-dimensional features. Experimental results demonstrate its use in image recognition, remote sensing, computational biolo…
Quantum Neural Networks Mislead: Kernel Methods with Trainability Bottlenecks, Not Neural Nets
Quantum neural networks (VQCs) are variational quantum circuits that map data to quantum states via nonlinear embeddings followed by linear operations, mathematically equivalent to kernel methods rather than composable linear-nonlinear chains of classical neural networks. Training scales quadratical…
Quantum Machine Learning Pioneer Shares Serendipitous Entry and Field Evolution from Niche to Industry-Driven
Maria Schuld entered quantum machine learning serendipitously during her PhD after failing to secure jobs in political science, leveraging freedom from low expectations to explore quantum walks in neural networks amid the field's nascent emergence. The field shifted from quantum computing researcher…
Quantum Circuits as Hybrid ML Models: Composable Like Neural Nets, Analyzable Like Kernel Methods
Variational quantum circuits function as machine learning models by encoding classical data into quantum states and performing trainable measurements, enabling composability and differentiability for seamless integration into PyTorch/TensorFlow pipelines via frameworks like PennyLane. This approach …
Quantum Circuits Enable Trainable Kernel Methods for Near-Term Machine Learning
Quantum machine learning replaces classical models in ML pipelines with parameterized quantum circuits, which are trained via parameter-shift rules for gradient computation and integrate seamlessly with frameworks like PyTorch and PennyLane. These circuits encode classical data into high-dimensional…
PennyLane Enables Differentiable Quantum Programming for Near-Term Hybrid Quantum-Classical Optimization
PennyLane integrates quantum circuits into automatic differentiation frameworks like JAX, PyTorch, and TensorFlow by providing gradients via parameter-shift rules, enabling gradient-based optimization of parameterized quantum circuits. This supports near-term noisy quantum devices through variationa…
Supervised Quantum ML Models Equate to Kernel Methods via Hilbert Space Inner Products
Supervised quantum machine learning models, often mislabeled as quantum neural networks, are mathematically kernel methods that operate in high-dimensional Hilbert spaces accessible only through measurement-derived inner products. These models can be reframed as support vector machines with kernels …
Quantum Annealing RBM Matches Classical Training for Cybersecurity Data Balancing and Classification
Researchers trained restricted Boltzmann machines (RBMs) using D-Wave 2000Q quantum annealing (QA) and classical contrastive divergence (CD) on the imbalanced ISCX cybersecurity dataset. Two balancing schemes—undersampling with majority voting and synthetic data generation—improved classification ac…
Data Encoding Strategy Controls Expressive Power of Quantum Machine Learning Models via Fourier Spectra
Parametrized quantum circuits for supervised learning act as partial Fourier series approximators, with accessible frequencies dictated by data encoding gates. Repeating simple encoding gates expands the frequency spectrum, enabling richer approximations. Models realizing all Fourier coefficients be…
Quantum Machine Learning Prioritizes Generalization Over Speedups via Trainable Circuits and Data Encoding
Quantum machine learning (QML) intersects quantum computing and ML, distinguishing classical-for-quantum, quantum-for-classical, quantum data with quantum processing, and quantum devices for classical data processing. Near-term QML treats quantum circuits as trainable parameterized models, emphasizi…
Data Encoding Dominates Quantum Machine Learning Success
Quantum machine learning with variational circuits encodes classical data into exponentially large Hilbert spaces as a feature map, making this step 95% of the challenge since subsequent operations are linear transformations followed by quadratic measurement nonlinearity. Optimal encodings cluster s…
Quantum Metric Learning Optimizes Embeddings for Analytic Measurements in Quantum Classifiers
Quantum classifiers consist of a feature map embedding classical data into Hilbert space followed by a trainable measurement. The proposed quantum metric learning trains the embedding to maximally separate classes under a specific metric (l1/trace or l2/Hilbert-Schmidt), making the optimal measureme…
Quantum Kernel Methods Leverage Hilbert Space Equivalence to Boost Classical ML Classifiers
Kernel methods in ML define similarity measures between data points via positive semi-definite functions, equivalent to inner products in a high-dimensional Hilbert feature space, enabling nonlinear classification through the kernel trick in algorithms like SVMs and Gaussian processes. Quantum theor…
Hybrid Quantum Transfer Learning Augments Pre-Trained Classical Networks with Quantum Circuits for Efficient NISQ Processing
The paper extends transfer learning to hybrid classical-quantum neural networks by pre-training classical networks on high-dimensional data like images and augmenting them with final variational quantum circuits. This leverages classical preprocessing to embed informative features into quantum proce…
Universal Continuous-Variable Quantum Neural Networks via Layered Gaussian and Non-Gaussian Gates
The paper presents variational quantum circuits in the continuous-variable (CV) architecture as universal neural networks, using layered Gaussian gates for affine transformations and non-Gaussian gates for nonlinear activations. These CV-QNNs encode highly nonlinear transformations unitarily, suppor…
Low-Depth Variational Quantum Classifiers with Poly-Log Parameters for Noisy Quantum Hardware
Proposes a variational quantum classifier encoding input features into quantum state amplitudes, processed by a shallow circuit of parameterized single- and two-qubit gates followed by single-qubit measurement. The architecture scales learnable parameters poly-logarithmically with input dimension, e…
Quantum Kernel Methods Mirror Feature Maps to Hilbert Spaces
Quantum computing parallels kernel methods by enabling efficient computation in high-dimensional Hilbert spaces via nonlinear feature maps that encode classical data into quantum states. Two quantum ML approaches emerge: (1) quantum estimation of intractable kernel inner products for classical algor…
Quantum Ensembles Enable Exponentially Large Untrained Classifier Aggregates via Parallel Evaluation
Quantum ensembles of quantum classifiers form by preparing a superposition state that encodes multiple classifiers, allowing parallel evaluation on a quantum computer followed by a single-qubit measurement for the collective decision. This approach supports exponentially large ensembles without indi…
Quantum Linear Regression Enables Direct Prediction via Single Qubit Measurement
The algorithm performs least-squares linear regression on quantum computers, prioritizing prediction of outputs for new inputs over parameter readout. It handles non-sparse data via low-rank approximations, reducing condition number dependence. Runtime is logarithmic in input dimension when data is …
Quantum Algorithm Leverages Hamming Distance for Enhanced Pattern Classification
This paper introduces a quantum machine learning approach to pattern classification by adapting Trugenberger's quantum Hamming distance measurement method. The algorithm aims to exploit quantum computing's known advantages over classical methods for specific tasks within machine learning. It demonst…
Quantum Perceptron via Phase Estimation Mimics Classical Step Activation with Linear Resources
The paper proposes a quantum perceptron model that simulates the step-activation function of classical perceptrons using the quantum phase estimation algorithm. It processes inputs of size n with O(n) resource requirements, enabling efficient scaling. This foundational unit supports development of t…
Systematic Overview of Quantum Machine Learning Approaches
Quantum machine learning explores enhancing classical ML algorithms using quantum computing, from efficient execution of costly subroutines to reformulating stochastic methods in quantum terms. The field addresses tasks like image/speech recognition and optimization relevant to IT. This paper provid…
No Existing Quantum Neural Network Fully Merges Neural Nonlinearity with Quantum Unitary Dynamics
Quantum Neural Networks (QNNs) aim to fuse neural computing's nonlinear, dissipative dynamics with quantum computing's linear, unitary evolution, but current proposals fall short. The paper systematically reviews QNN research, defines key requirements, and finds no model fully leverages both quantum…
Quantum Walks Model Associative Memory in Qubit-Based Neural Networks
Proposes quantum neural networks (QNNs) using qubits instead of binary neurons to leverage quantum computing. Models QNN dynamics via stochastic quantum walks on global firing state graphs, replicating classical associative memory. Biased discrete Hadamard walks from biological neuron updates fail u…