Figure 2
From: Experimental kernel-based quantum machine learning in finite feature space
Training results on a random inseparable data set of 40 samples (up/down-tipped triangles). The performance on a test set (left/right-tipped triangles) of 60 points (the fraction of correctly classified samples that were not used in the QML process) is given in the bottom right corner of each respective subplot. We find that the optimal variance/resolution choice for the Gaussian kernel is \(s=2\). For \(s=3\) we deal with overfitting. Shown are the simulation results both for an exact Gaussian kernel and for the truncated FM (8) comprising 4 terms (\(q=2\)). The learned classification boundaries are given as contour plots. The slight difference in performance compared to the theoretical prediction is due to statistical fluctuations in the experimental data and the relatively small test set (misclassification of a single near-boundary point results in a 0.02 performance drop).
