search

hongxin001 / logitnorm_ood

Official code for ICML 2022: Mitigating Neural Network Overconfidence with Logit Normalization
141 stars 14 forks source link

What's difference between "Unif" and "Original" in Figure 4(a) #7

Open zh-jp opened 3 months ago

zh-jp commented 3 months ago

Thank you for your work!

Could you give a further explanation about the difference between uniform distribution (Unif) and the original class priors of the training dataset. If the dataset used is cifar, I think the two should be the same.

Hope for your reply!

hongxin001 commented 3 months ago

I think you mean the work of open-sampling? In that work, we study the problem of data imbalance so the original prior is not uniform.

zh-jp commented 3 months ago

Oh, I just notice this repo isn't "Open-Sampling: Exploring Out-of-Distribution Data for Re-balancing Long-tailed Datasets", but I came in from the url provided of this paper.

Thank you for your reply.

And could you give the detail or example about the original prior, I still don't understand.

zh-jp commented 3 months ago

My understanding is that in the uniform distribution and the original, each sample has the same probability of being sampled.