view article
Figure 9
Reliability of the test data set classification, as shown by the confidence level represented by the CNN classification vector, for CNNs trained on data from the same LCLS run. Within each test data set of 600 images the left column shows classification probabilities for images correctly classified as Miss, while the right column shows the sum of the Hit + Maybe probabilities for those images correctly classified as either Hit or Maybe. Each panel gives the total number of correct classifications, and the average confidence probability for those images correctly classified. Notably, the L498 network, that has a low success rate for predicting Hit/Maybe (Table 3[link]), also has a low confidence level in Hit/Maybe classification. In contrast, for networks such as LN83, the probabilities associated with most Hit/Maybe images are above 90%, indicating that these images are relatively easier to classify.

Journal logoJOURNAL OF
SYNCHROTRON
RADIATION
ISSN: 1600-5775
Volume 25| Part 3| May 2018| Pages 655-670
https://doi.org/10.1107/S1600577518004873
Follow J. Synchrotron Rad.
Sign up for e-alerts
E-alerts
Follow J. Synchrotron Rad. on Twitter
Twitter
Follow us on facebook
Facebook
Sign up for RSS feeds
RSS