Speaker
Description
Kernel methods are used in Machine Learning to enrich the feature
representation of data in order to improve the generalization property
of models. Quantum kernels efficiently implement complex transformations
encoding classical data in the Hilbert space of a quantum system.
Working with quantum kernels may produce benefits with respect to
classical models. However, the choice of a best-performing quantum
embedding requires prior knowledge of the data which is typically not
available. We propose an algorithm that automatically selects the best
quantum embedding through a combinatorial optimization procedure
that operates on the structure of the circuit representing the quantum
embedding. The algorithm modifies the generators of the gates, their
angles (which depend on the data points), and the qubits on which
the various gates act in search of the best structure. Since combinatorial
optimization is computationally expensive, we have introduced a
criterion based on the exponential concentration of kernel matrix coefficients
around the mean to immediately discard an arbitrary portion
of solutions that may perform poorly. Contrary to the gradient-based
optimization (e.g. trainable quantum kernels), our approach is by construction
immune from the barren plateau problem. We demonstrate the
increased performance of our approach, with respect to randomly generated
quantum embeddings, on both artificial and real-world datasets.
A more detailed description of this algorithm, the experiments and their
results can be found in:
Incudini, M., Martini, F., Di Pierro, A.: Structure Learning of Quantum
Embeddings. arXiv (2022). https://doi.org/10.48550/ARXIV.2209.11144.
https://arxiv.org/abs/2209.11144
Email Address of submitter
massimiliano.incudini@univr.it
Short summary of your poster content
The Structure Learning algorithm automatically identifies a satisfactory quantum circuit to be used in machine learning tasks. The algorithm is based on a combinatorial optimization process; numerical properties unique to kernel methods (e.g. concentration of kernel coefficients around an average value) are used to speed up the optimization process.
| Poster printing | Yes |
|---|
