Uncertainty quantifcation in vision based classifcation

The past decade of artifcial intelligence and deep learning has made tremendous progress in highly perceptive tasks such as image recognition. Deep learning algorithms map high dimensional complex representations to low dimensional array mappings. However, these mappings are generally blindly assumed to be correct, further justifed with high accuracies on trending datasets. The challenge of creating a comprehensive, explainable and reasonable deep learning system is yet to be solved. One way to deal with this is by using uncertainty quantifcation, or uncertainty aware learning, with the help of Bayesian methods. This thesis contributes to the feld of uncertainty aware learning by demonstrating how uncertainty can be used to recover performance in case of a physical attack, how uncertainty can be used to improve sensitivity to noise and how it can be used to improve performance on dynamic datasets. The frst contribution involves learning from model uncertainty in the application of deep learning-based semantic segmentation. The second contribution deals with robustness and sensitivity analysis in image classifcation and fnally, the third contribution in continual learning by using variance to update the learning rate. The frst contribution proposes the architecture AdvSegNet which aims to improve the performance of Bayesian SegNet. In the second contribution, a combined architecture of convolutional network feature extractor and a Gaussian process (CNN-GP) is made to classify images under uncertain conditions including noise, blurring and adversarial attacks. Finally, in the continual learning subject area, the architecture CNN-GP is trained on datasets presented sequentially. Results show an improvement in performance and sensitivity to adversarial attack and noisy conditions as well as an improvement in dynamic datasets with a small number of tasks.