view article

Figure 11
Stochastic neighbor embedding of the feature maps extracted from an LG36-trained CNN applied to the LG36 dataset. Each spot on the map corresponds to one image, and is colored by its ground-truth classification of Hit (red), Maybe (blue) or Miss (green). On such an embedded map the x,  y coordinates have no specific physical interpretation; indeed, the final coordinates of the spots are dependent on the random number seed utilized. Rather, the utility of the plot lies in the fact that near neighbors within it display similar image characteristics, as encoded by their feature vectors. A visual inspection of the images then reveals what the similarities are. Image insets illustrate the cluster of Bragg-spot hits exemplified by shot 768, as well as the lower-right cluster, exemplified by shot 438, which turns out to contain no diffracted photons (no X-ray beam, dark noise only). Miss images that contain only a water ring, exemplified by shot 1161, form a continuous distribution, shown by the black arrows, with the arrows' arc roughly indicating an ordering from strongest to weakest water signal.

Journal logoJOURNAL OF
SYNCHROTRON
RADIATION
ISSN: 1600-5775
Follow J. Synchrotron Rad.
Sign up for e-alerts
Follow J. Synchrotron Rad. on Twitter
Follow us on facebook
Sign up for RSS feeds