My-implementation-of-What-Uncertainties-Do-We-Need-in-Bayesian-Deep-Learning-for-Computer-Vision

This is my implementation of classification task in paper What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision.

In this repo, I model aleatoric uncertainty, epistemic uncertainty and both of them of a classification task with MNIST dataset. I also implement a vanilla network to be compared with.

However, these are all based on my understanding of this paper. In regression task, it is easy to compute var(mean) (i.e. epistemic uncertainty) and mean(var) (i.e. aleatoric uncertainty), but in classification task, I really don't know how to compute aleatoric uncertainty and epistemic uncertainty for each sample, since they are all vectors rather than a single value. If I can compute them, I also don't know how to plot these uncertainties like regression task. (New: After reading so many papers, I think the best way to quantify uncertainties in classification task is entropy, though it is not comparable with the measurement of variance in regression task. In this situation, the sigma vector is just used in the loss function to mitigate the influence of noisy samples. Explicit quantification of uncertainties is better done by the entropy of mu. I'll implement this later.)

~~Besides, I still can't make sure that whether each logit value is drawn from a Gaussian and the whole logit vector is drawn from a multi-dimensional Gaussian distribution. I saw other repos predict variance of each sample by a single value, while I think it should be a vector hence the variance is a 'diagonal matrix with one element for each logit value'.~~ (This is discussed in issue #1)

The results seem that modeling aleatoric uncertainty can improve model's performance.

Any feedback and discussion is welcome.