Kaleidophon
commented
4 months ago Hey, thank you for reaching out!
I find this question a bit difficult to answer questions about this particular instance without knowing anything about the model, training data and training objective.
I can tell you that in general, the uncertainties proposed by Sensoy et al. are idealized - when you see these pretty plots of the Dirichlet distribution on the probability simplex in different situations, this is how we would like it to look, but that does not mean it is well-behaved in practice. You can imagine that simply because we predict a Dirichlet instead of a categorical using our neural network, does not mean that automatically everything would work out of the box. We need to train the network to display this behavior.
There are different strategies to achieve this though that you could try here: One is through the use of OOD examples (see the Malinin & Gales papers) to regularize the network, where you could either use different datasets, create OOD samples yourself by adding noise etc. to your input images, or even use an additional generative model to create OOD images for you. The other one is through inductive biases in the architecture, specifically the posterior network and natural posterior network by Charpentier et al., where you train a normalizing flow model on the latent representation of the underlying classifier. This helps to identify uncertain samples because their latent representations would (ideally) be assigned low probabilities under the flow model.
Hope that gave some useful pointers!
haruishi43
Your survey has helped me tremendously in understanding the landscape of EDL. I've recently started learning about uncertainty estimation for deep learning and wanted to experiment with EDL for dense classification tasks like semantic segmentation.
I started with EDL [Sensoy+ 2018] and visualized the epistemic uncertainty as well as the aleatoric and distributional uncertainties using the equations in your paper. I've noticed that all uncertainties have the same characteristics with different values, as shown in the figure below. I would think aleatoric uncertainties are higher for boundaries, while epistemic uncertainties are higher for OOD pixels/objects.
Do you happen to know why this is? My guess is that Sensoy's EDL is not particularly suited to produce uncertainties that can discriminate aleatoric and epistemic uncertainties.