Fig. 4: Evolution of GP parameters during training.
From: Imaging mechanism for hyperspectral scanning probe microscopy via Gaussian process modelling
Evolution of kernel length scales (a–d) and inducing points (e–h) during the stochastic variational inference (SVI)-based model training for full data set (a, e), and data corrupted by removing 70% of observations (b, f), 90% of observations (c, g), and 99% of observations (d, h). a The first two dimensions in a–d (dim 1 and dim 2) correspond to x and y coordinates, whereas the third dimension (dim 3) corresponds to frequency. The kernel length scales define the spatial resolution of the technique (assuming atomically thin domain wall width) in the spatial domain and the width of the resonance peak in the frequency domain.
