Uncertainty quantification has received increasing attention in machine learning in the recent past. In particular, a distinction between aleatoric and epistemic uncertainty has been found useful in this regard. The latter refers to the learner's (lack of) knowledge and appears to be especially difficult to measure and quantify. In this paper, we analyse a recent proposal based on the idea of a second-order learner, which yields predictions in the form of distributions over probability distributions. While standard (first-order) learners can be trained to predict accurate probabilities, namely by minimising suitable loss functions on sample data, we show that loss minimisation does not work for second-order predictors: The loss functions proposed for inducing such predictors do not incentivise the learner to represent its epistemic uncertainty in a faithful way.

Neural Information Processing Systems http://nips.cc/

Neural Information Processing Systems http://nips.cc/

Neural Information Processing Systems http://nips.cc/

Arya et al. suggest using either supervised or self-supervised learning techniques to train the NN; the former requires access to exact inference

Based on this, we believe that other classic9 methods would perform similarly to DeViSe if used instead in our GZSL semantic segmentation baseline. Yet, as abundantly exemplified in fully supervised learning, moving from image-level categorization to34 pixel-level recognition is not as direct or straightforward as it might seem. Moreover,toencode spatial context,37 we propose a novel graph convolutional generator which, conditioned on context graphs, generates corresponding38 structured pixel-level features. Also, as we shall clarify, our framework is not solely bound to GMMN as in [7]; it39 is in fact agnostic to the choice of the generative model.