Motivation of the method

[hathawayxxh]

Motivation of the method

[Kthyeon]

Hello,

thanks for your interest. We agree that class prototype and our methods are quite similar under clean data settings. However, we think that prototype can be more contaminated towards label noise while ours does not. Furthermore, with the use of eigenvector, the perturbation is more correctly estimated. Your question looks great discussion topic for further research.

[hathawayxxh]

Yes, I understand that the first eigenvector might be less contaminated by label noise. However, in the scenario of high noise_rate (e.g., 90% in the CIFAR10 dataset), it means the ratio of clean data in each class is quite small, will this avoids the eigenvector from represeanting the real distribution of the class?

[Kthyeon]

That is a good point. In the pilot experiment, we find that eigenvectors are more robust than other prototype-based methods such as prototype, anchor generated from Mahalanobis distribution, or Minimum covariance detector (MCD) method.

In my thought, eigendecomposition seems quite robust towards noisy representations compared to prototype which is based on the concept of simple averaging

[hathawayxxh]

Thanks, I totally understand that the eigenvectors are more robust than prototype-based methods. However, in my experiments, I find the FINE sampling method performs much worse than the simplest small-loss methods. Have you met this delimma in your experiments?

[Kthyeon]

In our experiments, we have not met such situations before...., but can be dependent on the kinds of datasets.

The benefits of eigenvectors can be affected by the ability of backbone network. So, how about using warmup stage for the early training with significant magnitude of weight decay (or other regularization)? and then how about applying the FINE method with such pretrained arch?