Adding Probabilistic Certainty to Improve Performance of Convolutional Neural Networks

Abstract

Convolutional Neural Networks (CNN) are successfully being used for different computer vision tasks, from labeling cancerous cells in medical images to identify traffic signals in self-driving cars. Supervised CNN classify raw input data according to the patterns learned from an input training set. This set is typically obtained by manually labeling the image which can lead to uncertainties in the data. The level of expertise of the professionals labeling the training set sometimes varies widely or some of the images used may not be clear and are difficult to label. This leads to data sets with pictures labeled differently by different experts or uncertainty in the experts opinions.

These kind of errors on the training set do happen more frequently when the CNN task is to classify numerous labels with similar characteristics. For example, when labeling damages on civil infrastructures after an earthquake, there are more than two hundred different labels with some of them similar to each other and the experts labeling the sets frequently disagree on which one to use. In this paper, we use probabilistic analysis to evaluate both the likelihood of the labels in the training set (produced by the CNN) and the likelihood’s uncertainty. The uncertainty in the likelihood is represented by a probability density and represents a spreading (as it were) of the CNN’s likelihood estimate over a range of values dictated by the uncertainty in the truth set.