Synopsis
CERN has started its Quantum Technology Initiative in order to investigate the use of quantum technologies in High Energy Physics (HEP). A three-year roadmap and research programme has been defined in collaboration with the HEP and quantum-technology research communities. In this context, initial pilot projects have been set up at CERN in collaboration with other HEP institutes worldwide on Quantum Computing and Quantum Machine Learning in particular. This talk will provide an overview of recent results obtained by the different studies, focusing on current usage of quantum machine learning techniques for HEP use cases and beyond.
From this talk:
- Alternative classical clustering methods to benchmark quantum clustering against (see the scikit-learn comparison):
- Centroid based: k-means and mini-batch k-means, which assume convex, roughly isotropic clusters of similar size and need the number of clusters.
- Mean shift and affinity propagation, which choose the number of clusters themselves, at quadratic or worse cost in the number of samples.
- Spectral clustering, which clusters the leading eigenvectors of a graph Laplacian built from a similarity kernel, and so handles non-convex shapes; this is the natural slot for a quantum kernel.
- Hierarchical: Ward and agglomerative linkage, and BIRCH for large datasets.
- Density based: DBSCAN, HDBSCAN and OPTICS, which find arbitrarily shaped clusters and label outliers as noise.
- Gaussian mixture models, fitted by expectation maximisation, which give soft assignments.
- Threshold decisions and metrics for a binary classifier with score and decision rule :
- Precision , recall (true positive rate, signal efficiency) , false positive rate (its complement is the background rejection).
- The ROC curve plots the true positive rate against the false positive rate as varies. Its area under the curve (AUC) is threshold independent and equals the probability that a random positive scores above a random negative; is random guessing.
- For imbalanced classes, common in HEP where signal is rare, the precision recall curve and its average precision are more informative than the ROC AUC.
- The threshold itself is then chosen by the use case: maximising Youden’s , maximising (the harmonic mean of precision and recall), or fixing a target signal efficiency or background rate.
- CR gate: in the Quantum Inspire knowledge base it is the controlled phase rotation , which reduces to CZ at . Not to be confused with the cross-resonance gate of fixed frequency transmons (IBM), also abbreviated CR, where the control qubit is driven at the target’s frequency, giving an effective interaction ; at plus single qubit rotations it yields a CNOT.
- Bell inequalities No local hidden variables model can represent Quantum Mechanics results.
- McNemar-Bowker test of symmetry (NCSS notes): for paired categorical outcomes in a contingency table (for example the labels given by a classical and a quantum classifier to the same test samples), it tests for all with
which is asymptotically distributed with degrees of freedom (minus one for each pair with ). For it reduces to McNemar’s test. It is the right test when the two models are evaluated on the same samples, unlike tests assuming independent samples.
- Tüysüz, Chang, Demidik, Jansen, Vallecorsa and Grossi, “Symmetry breaking in geometric quantum machine learning in the presence of noise”, arXiv:2401.10293: equivariant quantum neural networks can keep their equivariance under Pauli channels, but not under amplitude damping; the symmetry breaking grows linearly with the number of layers and with the noise strength. Supported by simulations and hardware runs up to 64 qubits, with proposed strategies to protect the symmetry.
- Geometric quantum machine learning encodes a symmetry group of the data, acting on the encoded states through a unitary representation , directly in the model:
- We want an invariant model, . This holds if the encoding is equivariant, , every layer commutes with the representation, , and the measured observable commutes with it too.
- Commuting generators can be built by twirling, (a Haar integral for a compact Lie group), which projects any generator onto the commutant of the representation.
- Restricting to the commutant shrinks the hypothesis space, which improves generalisation from few samples and can mitigate barren plateaus. As the paper above shows, hardware noise that does not commute with breaks this symmetry.