Bayesian Deep Neural Networks with Agnostophilic Approaches

A vital area of AI is the ability of a model to recognise the limits of its knowledge and flag when presented with something unclassifiable instead of making incorrect predictions. It has often been claimed that probabilistic networks, particularly Bayesian neural networks, are unsuited to this problem due to unknown data, meaning that the denominator in Bayes’ equation would be incalculable. This study challenges this view, approaching the task as a blended problem, by considering unknowns to be highly corrupted data, and creating adequate working spaces and generalizations. The core of this method lies in structuring the network in such a manner as to target the high and low confidence levels of the predictions. Instead of simply adjusting for low confidence, developing a consistent gap in the confidence in class predictions between known image types and unseen, unclassifiable data new datapoints can be accurately identified and unknown inputs flagged accordingly through averaged thresholding. In this way, the model is also self-reflecting, using the uncertainties for all data rather than just the unknown subsections in order to determine the limits of its knowledge. The results show that these models are capable of strong performance on a variety of image datasets, with levels of accuracy, recall, and prediction gap consistency across a range of openness levels similar to those achieved using traditional methods.