Evaluating Loss Functions for QNet: A Comparative Analysis on Modern NISQ Simulators

Abstract

Quantum Neural Networks (QNNs) hold substantial potential for advancing machine learning tasks on quantum computing systems. Using the most recent IBM quantum simulators, this paper presents QNet, a quantum neural network architecture designed for noise resilience. Our method involves switching from cross-entropy loss to focal loss, updating outdated Qiskit functions, and migrating the original setup to the FakeManilaV2 backend. The Epileptic Seizure Recognition dataset (UCI/Kaggle, 11500 samples) was used in the experiments. A subset of 2000 samples was used for noiseless simulations (178 features reduced by PCA to 8 dimensions over 50 epochs), and 500 samples for noisy simulations because of the computational limitations, over 10 epochs. To evaluate scalability and robustness, we later expanded our evaluation to the entire dataset. Results show that focal loss improves noise resilience and achieves faster convergence, highlighting its potential for useful QNN implementations on existing NISQ devices.