Semi-Weak Supervised Learning

Research Paper: arxiv.org/pdf/...
Blog Post: / billion-scale-semi-sup...
Github Repository: github.com/fac...
Facebook has recently presented a really interesting framework to make use of 1 billion weakly-supervised Instagram images (labeled with hashtags) using a model-distillation pipeline. Thanks for watching, please subscribe!

[@PeterOtt]

This paper is on my to-read list and now I'll understand what I'm getting into a whole lot more easily. Thanks for the great video!

[@connor-shorten]

Thank you so much! I really hope you find the paper interesting, I was surprised at the relative lack of interest following their first paper on weak supervision with instagram pictures!

[@PeterOtt]

@@connor-shorten I haven't read that either, guess that's going on the never ending list as well lol

[@nicolasessipova4376]

Henry! Thank you!!! 👏👏👏👏👏👏👏👏👏👏👏👏

[@connor-shorten]

Really hope you like this video! I thought the consideration of class imbalance for weakly supervised dataset and the need for inference accelerators for large scale model distillation were really interesting!

[@indraneilpaul1309]

Small Correction: The intro conflates semi-supervision and weak-supervision

[@blakeedwards3582]

Thank you so much! Great review.

[@connor-shorten]

Thank you!!

[@Clutchuhh21]

God bless, this was a great video.

[@Intu11110]

So what would be the point if you already have a model that does well in the first place, there is no need for the student model

[@saichandubobbili6115]

The student model gets much more better at classification, the student model runs from both teacher(itself, or a large capacity model) and labelled dataset.