Machine Learning for Planetary Science by Helbert Joern;D'Amore Mario;A

Author:Helbert, Joern;D'Amore, Mario;A
Language: eng
Format: epub
Publisher: Elsevier
Published: 2022-03-22T00:00:00+00:00

Figure 5.10 HiRISENet filtered performance: only examples with posterior probability (confidence) greater than or equal to a given threshold (x-axis) are included. The user can specify this threshold. Circles indicate performance with a confidence threshold of 0.9.

5.5.3 Model calibration and performance

Modern convolutional neural networks are often poorly calibrated, which means that their self-reported posterior probabilities do not reliably reflect empirical probabilities [16]. To improve the reliability of posterior probabilities, we explored various model calibration methods and found that the most effective method for this data set is temperature scaling [16], which estimates a scaling parameter (“temperature”) to adjust the neural network's logits prior to their conversion to probabilities. Reliability diagrams for the HiRISENet model before and after temperature scaling method are shown in Fig. 5.11. The expected calibration error (ECE) (average weighted error across the probability bins) and maximum calibration error (MCE) (maximum error across the probability bins) were improved (reduced) after temperature scaling method was applied.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.