An Optimized Uncertainty-Aware Training Framework for Neural Networks.

Uncertainty quantification (UQ) for predictions generated by neural networks (NNs) is of vital importance in safety-critical applications. An ideal model is supposed to generate low uncertainty for correct predictions and high uncertainty for incorrect predictions. The main focus of state-of-the-art training algorithms is to optimize the NN parameters to improve the accuracy-related metrics. Training based on uncertainty metrics has been fully ignored or overlooked in the literature. This article introduces a novel uncertainty-aware training algorithm for classification tasks. A novel predictive uncertainty estimate-based objective function is defined and optimized using the stochastic gradient descent method. This new multiobjective loss function covers both accuracy and uncertainty accuracy (UA) simultaneously during training. The performance of the proposed training framework is compared from different aspects with other UQ techniques for different benchmarks. The obtained results demonstrate the effectiveness of the proposed framework for developing the NN models capable of generating reliable uncertainty estimates.

Uncertainty-aware skin cancer detection: The element of doubt.

Artificial intelligence (AI)-based medical diagnosis has received huge attention due to its potential to improve and accelerate the decision-making process at the patient level in a range of healthcare settings. Despite the recent signs of progress in this field, reliable quantification and proper communication of predictive uncertainties have been fully or partially overlooked in the existing literature on AI applications for medical diagnosis. This paper studies the automatic diagnosis of skin cancer using dermatologist spot images. Three different uncertainty-aware training algorithms (MC dropout, Bayesian Ensembling, and Spectral Normalized Neural Gaussian Process) are utilized to detect skin cancer. The performances of the three above-mentioned algorithms are compared from different perspectives. In addition, some images from the Cifar10 dataset are applied as the out-of-domain data and the performances of the algorithms are evaluated and compared for images that are far from the training samples. The accuracy, uncertainty accuracy, uncertainty accuracy for out-of-domain distribution samples, and the uncertainties of the predictions are reported in all cases and compared.