What is Contrastive Learning? (Contrastive Learning/Self-supervised Learning Explained)
Вставка
- Опубліковано 20 тра 2024
- In this video, I give an overview of an important AI topic called Contrastive Learning. This is used in
the popular VQGAN+CLIP models that can learn to make generative art that has never existed!
If you want to watch other machine learning/computer science videos, subscribe!
Link to SimCLR paper: arxiv.org/pdf/2002.05709.pdf
Link to FB AI Blog: / self-supervised-learni...
0:00 Intro
0:34 Motivating Example
1:28 Contrastive Learning Framework
5:15 Fine Tuning
5:35 Outro
#contrastivelearning #AI #ML - Наука та технологія
Just what I was looking for. Super clear explanation accompanied by great visualisations. Thanks man!
You're welcome! Glad you enjoyed it
I didn't understand the collapse to constant concept till this video. Thanks! Much love
Awesome work, thank you sooooo much!
this video was very ambiguous for me, I don't understand too much from that but thank you for suggesting that two papers I read them and I hope I understand that. Thank you so much keep on with more videos
This video did power my heart
Nice and clear, thanx
For image classification, if no label is ever provided, how does the model know that the same crop belongs to "Hotdog"?
Thanks! Helped out a lot :)
Mega underrated- you earned yourself a sub :D
Thank you!
where do these fractal-like videos in the outro come from? do they have to do with contrastive learning? they are absolutely stunning
wow! Amazing video - great explanation, subbed! Btw, I wonder if there is anything they can do with the g() function that they drop?
This is one of the best 6 minute summaries on an AI model I’ve seen
Oh wow, thanks for the compliment!
cool vid mate
Why do we drop the g function? Don't we lose some representation structure with that? Wouldnt it be more efficient to train classification with both f and g functions?
Great question! The g function is used for the task of mapping two vectors to a scalar value, z_i and z_j in my example. However, the downstream task (e.g., classification) may require a whole vector of outputs (i.e., one output per class giving the probability of each class).
We are essentially doing transfer learning here, where you keep the rest of model the same, but change the output layer.