Speaker
Description
The underlying likelihood of a given event originating from a partonic-level process is known to be approximately invariant under the Lorentz group. We find that quantum neural networks equivariant under such continuous symmetries exhibit improved generalization, sample and training time complexity. We show that this property is induced by the number of distinct group orbits in the data, with an increasing separation as the number of training samples outgrows the number of orbits. From the conservation laws of the Lorentz group, we build a quantum neural network invariant under $(\eta, \phi)-$translations, and compare it against another ansatz without the same inductive bias on a quark-gluon tagging task, numerically confirming our findings.