2020/03/fixmatch-semi-supervised/ about blog-comments HOT 5 CLOSED

Deep thanks for this awesome illustration. But I have some inquiries:

• First if the threshold is not achieved, do they bypass this sample for example?
• Regarding curriculum learning, I don't think that condifence has to increase within time as the model can have predict wrong prediction confidently even if it is not trained? What the paper to did ensure that curriculum learning is work as expected?

from blog-comments.

Additionally, why do we trust the model predictions on the weak augmentations, what if the model predict the weak augmented version incorrectly?

from blog-comments.

Thank you for the great post.

Can you please elaborate on how the threshold was achieved?

Thank you

from blog-comments.

Hi Amit!

Very thoughtful post!

I was wondering if it is appropriate to select the threshold based on output of the softmax activations.

Consider the following thread on twitter.

If we were to think in this point, we should be carefully selecting logits instead of the confidence scores right? Because let's say in cat dog classification problem, the logits being [1,2] and [-2,-1] gives the same probability distribution but inherently confidence scores connotation should be different right? Refer the following image

https://drive.google.com/file/d/1ajf2DWITt1z9k5YmgxfR3UWLveeg4Atf/view?usp=sharing

Thanks & Regards,
Vinayak.

from blog-comments.

Would be interested to see use the model as encoder to create features then do unsupervised clustering . Think it may achieve similar accuracy.

from blog-comments.