Scalable Uncertainty for Computer Vision with Functional Variational Inference

As Deep Learning continues to yield successful applications in Computer\nVision, the ability to quantify all forms of uncertainty is a paramount\nrequirement for its safe and reliable deployment in the real-world. In this\nwork, we leverage the formulation of variational inference in function space,\nwhere we associate Gaussian Processes (GPs) to both Bayesian CNN priors and\nvariational family. Since GPs are fully determined by their mean and covariance\nfunctions, we are able to obtain predictive uncertainty estimates at the cost\nof a single forward pass through any chosen CNN architecture and for any\nsupervised learning task. By leveraging the structure of the induced covariance\nmatrices, we propose numerically efficient algorithms which enable fast\ntraining in the context of high-dimensional tasks such as depth estimation and\nsemantic segmentation. Additionally, we provide sufficient conditions for\nconstructing regression loss functions whose probabilistic counterparts are\ncompatible with aleatoric uncertainty quantification.\n