"Train classical, deploy quantum" requires rethinking generalization
A converged moment-matching loss does not tell you whether a quantum generative model generalizes. We prove why, and we measure it across thirteen models at up to 30 qubits.
Scalable Message-Passing Quantum Graph Neural Networks in the Weisfeiler-Leman Hierarchy
A quantum graph neural network that passes messages inside the circuit, is exactly permutation equivariant, and reaches a chosen level of the Weisfeiler-Leman hierarchy. We prove the three properties and test them in simulations of up to 56 qubits.
Adaptive Directional Gradients for Parameterised Quantum Circuits
A gradient estimator for quantum circuits that averages V random directional derivatives. SPSA, random coordinate descent, and the parameter-shift rule are its limiting cases, and the QUIVER optimiser sets V from the measurement statistics.
Bayesian Quantum Orthogonal Neural Networks for Anomaly Detection
We put Gaussian distributions on the angles of orthogonal quantum layers, build a 3D convolutional autoencoder from them, and use it to find defects in 3D-printed parts. Part of the pipeline runs on IBM Brisbane.
QuIC: Quantum-Inspired Compound Adapters for Parameter Efficient Fine-Tuning
An adapter built from one small orthogonal matrix and its compound matrices fine-tunes DeBERTa, DINOv2, and LLaMA with under 0.02% of the base model's memory, over 40 times smaller than LoRA.