Benchmarking loss functions for trainable quantum feature maps
arXiv:2607.12487
The paper evaluates how different loss functions affect the training of quantum feature maps used in quantum machine learning, finding that Log-Likelihood Loss offers stable optimization with linear computational cost.
Abstract
Many quantum machine learning models employ quantum feature maps to encode classical data into quantum states. While fixed feature maps often lack sufficient expressivity for complex nonlinear classification tasks, trainable quantum feature maps (TQFMs) enable adaptive quantum kernels with enhanced learning capability. Different loss functions can induce distinct optimization dynamics, yet their effects remain poorly understood. In this work, we apply the Log-Likelihood Loss function for TQFMs and provide a systematic comparison with Distance Loss and Measurement Loss. Through extensive numerical experiments, we compare their optimization dynamics, computational costs, and classification performance. Our results show that Log-Likelihood Loss consistently achieves more stable optimization than Measurement Loss while retaining linear computational complexity. The resulting benchmark offers practical guidance for balancing trainability, computational efficiency, and predictive performance in quantum kernel optimization.
10 pages, 9 figures