Understanding Generalization in Quantum Machine Learning with Margins
Authors: Tak Hur, Daniel K. Park
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experimental studies on the quantum phase recognition dataset demonstrate that margin-based metrics are strong predictors of generalization performance, outperforming traditional metrics like parameter count. By connecting this margin-based metric to quantum information theory, we demonstrate how to enhance the generalization performance of QML through a classical-quantum hybrid approach when applied to classical data. |
| Researcher Affiliation | Academia | 1Department of Statistics and Data Science, Yonsei University, Seoul, Republic of Korea 2Department of Applied Statistics, Yonsei University, Seoul, Republic of Korea. Correspondence to: Tak Hur <EMAIL>, Daniel K. Park <EMAIL>. |
| Pseudocode | No | The paper describes methods through mathematical formulations and textual descriptions of processes and algorithms, but it does not contain any explicitly labeled pseudocode blocks or algorithms with structured steps. |
| Open Source Code | No | The paper does not contain any explicit statement about releasing source code for the methodology described, nor does it provide a link to a code repository. It mentions using third-party tools like PennyLane but not its own implementation code. |
| Open Datasets | Yes | Figure 4 presents the classification of MNIST (Le Cun et al., 2010), Fashion-MNIST (Xiao et al., 2017), and Kuzushiji-MNIST (Clanuwat et al., 2018) datasets using 8-qubit QCNN with various quantum embedding schemes. |
| Dataset Splits | Yes | The model was trained on 20 data points, evenly split across four classes. A small training sample was deliberately chosen to explore the overfitting regime, where labels were intentionally randomized with noise, following a methodology similar to that of Gil-Fuster et al. (2024). The test accuracy was measured on 1,000 test samples far exceeding the size of the training set to provide a robust estimate of true accuracy. ... For the fixed embedding, we used the ZZ Feature Map with three repeated layers. ... Unlike previous experiments, we did not examine the overfitting regime under label noise, choosing instead to utilize the full training and test datasets. |
| Hardware Specification | No | The paper does not explicitly specify the hardware (e.g., GPU, CPU models, or cloud computing resources) used to run the experiments. |
| Software Dependencies | No | The paper mentions software like the Adam optimizer and PennyLane, but it does not provide specific version numbers for these or any other key software dependencies required for replication. |
| Experiment Setup | Yes | The model was trained using Adam optimizer, with a learning rate of 0.001 and full-batch updates (Kinga et al., 2015). The model was trained for up to 5,000 iterations, with early stopping triggered based on a convergence interval of 500 iterations. Specifically, the training halted when the difference between the average loss over two consecutive intervals became smaller than the standard deviation of the most recent interval. ... The experimental setup remained the same as before, except for using a batch size of 16 instead of full-batch training. |