Fig. 3.

Machine-learning-driven imaging results of four different cases in Fig. 2d. Different numbers of measurements, 100, 200, 400, and 600, were performed with linear embedding techniques of random projection and PCA. More detailed results are recorded in Supplementary Videos 3 and 4. This set of panels clearly demonstrates that the reconstructions by PCA have overwhelming advantages over those obtained by the random projection for small number of measurements, since a large amount of relevant training samples are incorporated in PCA