Bayesian Convolutional Neural Network with Prediction Smoothing and Adversarial Class Thresholds
Abstract
Using convolutional neural networks (CNNs) for image classification for each frame in a video is a very common technique. Unfortunately, CNNs are very brittle and have a tendency to be over confident in their predictions. This can lead to what we will refer to as "flickering," which is when the predictions between frames jump back and forth between classes. In this paper, new methods are proposed to combat these shortcomings. This paper utilizes a Bayesian CNN which allows for a distribution of outputs on each data point instead of just a point estimate. These distributions are then smoothed over multiple frames to generate a final distribution and classification which reduces flickering. Our technique is able to reduce flickering by 67 . We also propose a second method to combat False Positive predictions of certain adversarial classes, or classes that have some cost if predicted incorrectly. This is accomplished by increasing the confidence threshold the adversarial class must meet in order to be the final predicted class. This technique is able to reduce false positives by 5.43 , while maintaining accuracy.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 01, 2022
- Accession Number
- AD1172378
Entities
People
- Noah Miller
Organizations
- Air Force Institute of Technology