|
The overfitting of deep learning (DL) models is one of the problems in artificial intelligence (AI) systems. It occurs when a DL model learns not only the patterns inherent to the data but also learns the noise characteristics and random fluctuations in the data. This learning behavior and the uncertainty caused by the noisy observations can negatively affect the decisions and the explainers of an AI system. The decision-makers are hidden in the semantic meanings of the feature maps. Hence, locating the feature maps that support/oppose AI’s decisions under uncertainty, caused by the noise characteristics and random fluctuations in the data, is challenging. However, the Bayesian search theory (BST) that has been widely used in target-tracking under noisy conditions could provide us with a solution to this research problem. This paper studies and proposes a BSTbased approach that assumes prior knowledge of the feature maps of the noise-free observations and generates posterior probabilities to find correlated feature maps of the noisy observations. The posterior probabilities built on a Gaussian likelihood is used for extracting semantic meanings that support/oppose the decisions hidden in feature maps. Hence, it can provide post hoc explainers while an AI system makes its predictions by using the feature maps of the final convolutional layer. In our simulation, we have used a pretrained VGG16 model that consists of 512 channels (or feature maps), where each channel provides 196 (i.e., 14x14) semantic meanings, in its final convolutional layer to make predictions. We have also used two bird images (Blue Jay and Indigo Buntings), and added varying uncertainty using Gaussian noise with distinct noise factors (0, 4, .., 20). Simulations show that we can precisely locate the feature maps and their semantic meanings that support/oppose the decision (i.e., the prediction) of an AI system under uncertainty.
|
