Beyond fixed circuits
A quantum feature map encodes a classical input into a quantum state. A kernel then measures similarity between encoded inputs. The choice of feature map shapes the similarities available to a downstream classifier.
Search over structure
An evolutionary approach treats circuit structure as a candidate solution. Mutation changes a candidate, evaluation measures its quality, and selection determines the candidates retained for future generations. The representation defines the architectural changes the search can express.
The evaluation matters
A candidate that scores well on the search objective does not automatically generalise. Separating search, model selection and final testing helps distinguish improvements to the objective from improvements on unseen examples.
An open direction
The appealing idea is to automate part of the architectural choice while keeping the search cost manageable. Circuit expressivity, evaluation cost and generalisation remain connected design questions.