A conditional diffusion model reconstructs quantum dot charge stability diagrams from just 4% of measured data, outperforming classical interpolation—especially for line-cut scans—enabling up to 5× faster spin qubit device characterization.
#SpinQubits #QuantumML #Research
2-qubit QRNNs with CNOT gates learn entanglement-based memory distinct from classical strategies, confirmed by causal tests (p<0.0001, d=0.89), but degrade to chance on IBM hardware while classical geometric strategies survive perfectly.
#QuantumML #MechanisticInterpretability #Research
WSBD optimizer uses gradient-derived importance scores to freeze low-impact QNN parameters, cutting circuit evaluations by up to 63.9% vs Adam. Formally proven to converge, it outperforms SGD/Adam on VQE, MNIST & parity tasks — saving hours of real QPU time.
#QuantumML #QNN #Research
LQAS, an evolution-inspired heuristic, automates PQC design via localized gate modifications. Tested on synthetic & quantum chemistry datasets, it boosted R² from -0.993 to 0.958 and was deployed on IBM 156-qubit Heron r2 processors.
#QuantumML #QuantumCircuits #Research
GNN trained on SDRG achieves ~94% pairing accuracy for disordered long-range quantum spin chains, reproducing entanglement entropy across all subsystem sizes and extending to finite-temperature via SDRG-X without retraining.
#QuantumML #QuantumSimulation #Research
CSIRO study shows 8-qubit entangled PQCs outperform separable circuits as PPO feature extractors in competitive Pong, matching classical MLPs in low-parameter regimes. CKA analysis confirms entangled circuits learn structurally distinct representations.
#QuantumML #QuantumAdvantage #Research
ML identifies the thermodynamic arrow of time on a 10-qubit NV-center diamond processor. A CNN achieves ~92% accuracy, k-means ~91%, and a diffusion model reproduces entropy production and directional heat flow without prior physical knowledge.
#QuantumThermodynamics #QuantumML #Research
New unified framework separates barren plateau causes in quantum circuits: mid-circuit information loss and scrambling suppress gradients independently of observable concentration, even in QCNN-inspired architectures—reshaping trainability analysis for QML.
#QuantumML #BarrenPlateaus #Research
ReQGAN synthesises full images from a single D-qubit PQC by pairing a learnable Neural Noise Encoder with a differentiable Intensity Calibration module, slashing FID scores ~50% vs. PQWGAN on MNIST within tight NISQ qubit budgets.
#QuantumML #QGAN #NISQ
Hybrid quantum-classical GNN using parameterized quantum circuits with attention mechanisms achieves competitive intrusion detection on 4 benchmarks, validated on real IBM quantum hardware under NISQ noise conditions.
#QuantumML #CyberSecurity #Research
First quantum autoencoder deployed in a federated learning framework for IoT anomaly detection. Tested on real Raspberry Pi hardware, QFL matches centralized training (F1 up to 0.99) while preserving data privacy via local-only parameter sharing.
#QuantumML #FederatedLearning #Research
A quantum perceptron transmitting only a qubit achieves ~85% prediction accuracy vs 75% classical—a provable, universal advantage across all linearly separable Boolean functions, linking quantum information theory to statistical learning theory.
#QuantumML #QuantumAdvantage #Research
BUPT researchers developed L-QMVKL, fusing cross-view quantum kernels with local data structure via hybrid global-local kernel alignment. Evaluated on Mfeat dataset, it achieves significant accuracy gains competitive with classical ML models.
#QuantumML #QuantumKernels #News
Recurrent quantum circuits trained via a novel parameter-shift rule generate stochastic process samplers with linear circuit complexity. Outperform Born machines by orders of magnitude in accuracy, especially in data-sparse regimes.
#QuantumML #QuantumComputing #Research
Rigetti Computing partners with Moody's Analytics to pursue narrow quantum advantage in finance, targeting practical problems where quantum methods outperform classical computing in speed, cost, or capability.
#QuantumML #QuantumFinance #News
First PAC-Bayesian generalization bounds for quantum ML via layered quantum channels, including dissipative/mid-circuit operations. Non-uniform, data-dependent bounds validated on quantum phase classification—offering actionable model design insights.
#QuantumML #QuantumComputing #Research
Replacing standard binary with Reflected Gray Code in Quantum Circuit Born Machines improves training on numerical data with zero overhead, outperforming standard encoding in 82% of simulations by preserving data structure in qubit space.
#QuantumML #QCBM #Research
A hybrid QGAN with VAE latent-space encoding and style-based data re-uploading generates novel drug-like molecules on IBM's 156-qubit Heron chip using just 110 trainable parameters—6,400× fewer than classical GAN—with competitive MOSES benchmark scores.
#QuantumML #DrugDiscovery #Research
A quantum algorithm for the Random Forest regression testing phase achieves query complexity O(t·h) independent of tree count n, outperforming classical O(n·h), using Quantum Amplitude Estimation on encoded decision tree leaf values.
#QuantumML #QuantumAlgorithms #Research
New research quantifies how increasing symmetry in spin models (e.g., Ising→XXZ) exponentially amplifies quantum 'shots' needed for fidelity-based kernel estimation, delivering symmetry-aware bounds for quantum ML in materials science.
#QuantumML #QuantumPhaseTransitions #News
FCAT & Xanadu adapt the Hidden Subgroup Problem framework to handle noisy, real-world data—enabling quantum algorithms to find approximate patterns rather than requiring exact mathematical structures, expanding practical QML applicability.
#QuantumComputing #QuantumML #News
A silicon photonic QRC using a single 'kitten' state achieves ~97% classification accuracy—a 9-fold error reduction over classical inputs—simulated exactly via generalized-P phase-space methods with no Hilbert space cutoff.
#QuantumReservoirComputing #SiliconPhotonics #QuantumML
IBM Research & ETH Zurich built a 124-qubit QELM with 5,000+ two-qubit gates, overcoming noise & concentration effects via hyperparameter tuning & eigentask analysis, matching classical ML on time-series forecasting & satellite image classification.
#QuantumML #ReservoirComputing #News
Hybrid 4-qubit quantum circuit + Vision Transformer boosted flood detection accuracy from 84.48% to 94.47% (F1: 0.841→0.944) on SEN12-FLOOD satellite imagery, cutting false positives in non-flooded terrain classification.
#QuantumML #RemoteSensing #Research
QELM deployed on IBM Quantum using 124 qubits & 5,000+ two-qubit gates. Noise-robust hyperparameter tuning and local eigentask analysis yield performance competitive with classical baselines on time-series forecasting and satellite image classification.
#QuantumML #ReservoirComputing #Research
Novel LSTM + Quantum Circuit Born Machine hybrid outperforms classical LSTM on SSE & CSI 300 indices, achieving up to 66.67% MSE reduction via decoupled training that avoids data encoding bottlenecks and barren plateau issues.
#QuantumML #QuantumFinance #Research
Introduces GBBMs — photonic quantum generative models using parametrized Gaussian Boson Sampling circuits trained classically via MMD² loss. Scaled to 805 modes and 1.3M parameters on a single GPU in ~100 min, outperforming classical RBM baselines.
#QuantumML #GaussianBosonSampling #Research
QML models outperform classical neural networks against training data corruption, exhibiting phase transition-like noise resilience. A new quantum machine unlearning framework enables efficient removal of corrupted data influence—a key advantage for trustworthy AI.
#QuantumML #AISecurityl #News
WTHaar-Net replaces Hadamard mixing with Haar wavelets in hybrid quantum-classical CNNs, cutting MACs by 44% and surpassing ResNet on Tiny-ImageNet. Validated on IBM Quantum 127-qubit hardware with 4-qubit circuits.
#QuantumML #HybridQuantumClassical #DeepLearning
BMW/UBC researchers train end-to-end quantum Wasserstein GANs on full MNIST & Fashion-MNIST using just 11–13 qubits—no dimensionality reduction or patch tricks. Task-specific circuit design + multimodal noise achieves new QGAN state-of-the-art FID across all 10 classes.
#QuantumML #QGAN #Research
