Bayesian Augmentation of Convolutional Neural Network - Long Short Term Memory for Video Classification with Uncertainty Measures
[Technical Report, Master's Thesis]
AIR FORCE INSTITUTE OF TECHNOLOGY WRIGHT-PATTERSON AFB OH
Pagination or Media Count:
The success of Department of Defense DoD missions rely heavily on intelligence, surveillance, and reconnaissance ISR capabilities, which supply information about the activities and resources of an enemy or adversary. To secure this information, satellites and unmanned aircraft systems collect video data to be classified by either humans or machine learning networks. Traditional automated video classification methods lack measures of uncertainty, meaning the network is unable to identify those cases in which it predictions are made with significant uncertainty. This leads to misclassification, as the traditional network classifies each observation with same amount of certainty, no matter what the observation is. Bayesian neural networks offer a remedy to this issue by leveraging Bayesian inference to construct uncertainty measures for each prediction. Because exact Bayesian inference is typically intractable due to the large number of parameters in a neural network, Bayesian inference is approximated by utilizing dropout in a convolutional neural network.
- Computer Systems
- Unmanned Spacecraft