Improving Deep Network Robustness to Unknown Inputs with Objectosphere

Akshay Raj Dhamija, Manuel Gunther, Terrance E. Boult; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019, pp. 87-90

Deep Neural Networks trained on academic datasets often fail when applied to the real world. These failures generally arise from unknown inputs that are not of interest to the system. The mis-classification of these unknown inputs as one of the known classes highlights the need for more robust deep networks. The problem of identifying samples that are not of interest to the system has previously been tackled by either thresholding softmax, which by construction cannot return none of the known classes itself, or by learning new features for the unknown inputs using an additional back- ground or garbage class. As demonstrated, both of these approaches help but are generally insufficient when previously unseen classes are encountered. This paper overviews our recent publication Reducing Network Agnostophobia, NeurIPS 2018. The paper presented two novel loss functions that effectively handle unseen classes while providing a new measure for uncertainty. The ability to identify unknown samples plays a crucial role in developing robust networks that may be used in open-world problems. The paper also introduced an evaluation metric that focused on comparing performance of multiple approaches in an open-set setting.