Contrastive Clustering with SwAV
Ғылым және технология
This video explains a new algorithm combining self-supervised contrastive learning with clustering to learn better representations. This is done by predicting clusters from image features, and is extremely close in performance to full supervised learning, constrained on the ResNet50 architecture. Thanks for watching! Please Subscribe!
Paper Links:
Facebook AI Blog Post on this Paper: / high-performance-self-...
Contrasting Cluster Assignments: arxiv.org/abs/2006.09882
Salesforce Blog Prototypical Contrastive Learning: blog.einstein.ai/prototypical...
Prototypical Contrastive Learning Paper: arxiv.org/pdf/2005.04966.pdf
Supervised Contrastive Learning: arxiv.org/pdf/2004.11362.pdf
SimCLR: arxiv.org/pdf/2002.05709.pdf
Bootstrap your own Latent: arxiv.org/pdf/2006.07733.pdf
CURL: arxiv.org/pdf/2004.04136.pdf
Пікірлер: 18
2:10 Motivation Contrastive vs. Generative self-supervised learning 3:02 Contrastive Learning Progress 4:25 Instance Discrimination 7:42 Criticism of Instance Disrimination (I think these are really interesting issues raised) 9:02 Contrasive Instance Learning vs. Cluster Prediction 10:47 Contrast with Prototypical Contrastive Learning 11:42 Motivation of Online Clustering 14:16 Multi-Crop Augmentation 14:47 Results 16:36 I’d like to see more of this and less linear evaluation of representations (especially from big labs like Facebook)
Wow, mind-blowing. U keep generating high quality content faster than one can consume spanning many sub domains, great commitment and fantastic explanations.
@connorshorten6311
3 жыл бұрын
Thank you so much!
You are doing god's work mate.
@connorshorten6311
3 жыл бұрын
Lol, thank you!
@billykotsos4642
3 жыл бұрын
Lmao that profile pic
Thanks for the summary!
Been watching since day one Henry because of your unique and great coverage of SOTA topics. Keep it up and thank you!!!
@connorshorten6311
3 жыл бұрын
Thank you so much!!
Man you literally saved my presentation, again haha, thanks
@connorshorten6311
3 жыл бұрын
Glad to hear it!
Hello, I love your videos..a normal student trying to understand the latest and greatest research happening. I wanted to ask if these algorithms can be used for medical images?
@connorshorten6311
3 жыл бұрын
Definitely! See - "MoCo Pretraining Improves Representation and Transferability of Chest X-ray Models"! Goodluck with your research!
Can make vid of ALAE if u have interest on generative models? I'm still having hard time to understand this paper. It's first try to use latent space autoencoder and results are amazing
@connorshorten6311
3 жыл бұрын
Haha I was scared off by the double integral! I have a poor understanding of how they marginalize over latent variables for models like that, I've been working through Chapter 19 of Bengio, Goodfellow, and Courville's deep learning book to get a better sense of it. I think you might find it interesting as well, www.deeplearningbook.org/
I come here to find how is Q calculated, and learned it's in the appendix :\