Decision making under uncertainty with convolutional deep Gaussian processes

Deep learning models such as convolutional neural networks have brought advances in computer vision and were found to surpass human accuracy in computer vision problems. This has resulted in their use in many safety critical applications such as autonomous driving and healthcare, where decision making under uncertainty is crucial. However, deep learning models are vulnerable to out of sample and adversarial examples and they can be very risky to use in safety critical applications. Deep Gaussian process provide a Bayesian non-parametric approach to deep learning and are capable of modelling the uncertainty in data and model. In this paper, we show the uncertainty quantification capabilities of Convolutional deep Gaussian processes for computer vision problems. We show that convolutional deep Gaussian processes provide better uncertainty estimates for various settings such as out-of-distribution samples, adversarial attacks and uncertainty calibration experiments. They are found to be more robust than convolutional neural networks, Bayesian convolutional neural networks and deep Gaussian processes for image classification