search

danruod / IEDL

[ICML 2023] Offical implementation of the paper "Uncertainty Estimation by Fisher Information-based Evidential Deep Learning".
MIT License
31 stars 4 forks source link

A Question About Applying IEDL to Other Datasets? #2

Open YeungChiu opened 6 months ago

YeungChiu commented 6 months ago

Your work is very amazing and I have received a lot of inspiration on it, for which I am grateful.

Now I have a question, when I apply IEDL to other datasets, such as video dataset for video classification, I unexpectedly found that unlike image dataset, each video data and the $I(\alpha)$ obtained after the computation is overwhelmingly negative, however, the Fisher value that I have seen in the paper should be greater than or equal to zero.

I am confused about this, could you please give any relevant insights?

I would appreciate it.

danruod commented 1 month ago

Hello, sorry for the late reply. Thank you for your interest in our work. Do you mean that $- \log \mathcal{I}(\boldsymbol{\alpha})$ (corresponding to _loss_det_fisher__ in the code) is mostly negative? If so, you can print the values ​​of $\boldsymbol{\alpha}$, maybe they are too small. You can try to improve this situation by adjusting the coefficients of KL loss.