Self-supervised vision
Вставка
- Опубліковано 11 чер 2024
- An overview of the use of self-supervised learning in Computer Vision.
Timestamps:
00:00 - Self-supervised vision
00:29 - Self-supervised learning - motivation
01:39 - Self-supervised learning - motivation (cont.)
04:11 - Self-supervised learning - creating your own supervision
05:38 - Self-supervised learning - creating your own supervision (cont.)
06:32 - Self-supervised learning - creating your own supervision (cont. cont.)
07:36 - Self-supervised learning - context as supervision
09:09 - Back to vision: context as supervision
11:39 - Pretext task: inpainting
13:30 - Pretext task: jigsaw puzzles
14:32 - Pretext task: colourisation
16:00 - What's wrong with L2?
17:29 - Pretext task: counting
19:53 - Grouping/Common fate
20:36 - Pretext task: Grouping/Common fate
22:06 - Pretext task: Rotations
23:28 - Pretext task: Clustering
25:00 - Contrastive Learning
26:47 - Masked Autoencoders
Topics: #computervision #ai #introduction
Notes:
This lecture was given as part of the 2022/2023 4F12 course at the University of Cambridge.
It is an update to a previous lecture, which can be found here: • Self-supervised learni...
Links:
Slides (pdf): samuelalbanie.com/files/diges...
References for papers mentioned in the video can be found at
samuelalbanie.com/digests/2023...
For related content:
- Twitter: / samuelalbanie
- personal webpage: samuelalbanie.com/
- UA-cam: / @samuelalbanie1
The title of this video waaaaay understates how much information is packed into this!
This is your fifth or so video that I've watched, and you are the damn goat. 🙇
I love how you build up a dense slide that serves as review material for later followup/self-quizzing. You rock.
Hi, just want to say that I really like your presentation style in your videos. A lot of information (with historical context!), explained so calmly and clearly as I haven't seen it with any other AI-focused UA-camr before.
Thanks!
This channel is really underated, just continue man, just a matter of time before your videos will explode
From a highlevel understanding POV, I found the information provided at the first 10min of your video are widely spread, makes it hard to understand the topic and overall picture you are telling. You introduced the concepts of exploiting the redundant signal from mutli-modal, then history about SSL, then history about redundant signal and the sample storage needed for learning, then multimodal exploiting history, then history of SSL in NLP, then SSL in CV, after that pretext tasks until contrastive learning and MAE. Again, I understood the links and connected dots between them, but still very abroad and hard to catch.
Generally, I love your videos, good work and well done. May I suggest putting a conclusion that connects the dots after presenting each one of them, or putting a recap every few slides, to help paving the understanding path nicely?!
All the best
This is super interesting!
What's the difference between in painting and masked autoencoder? Seems like a transformer vs connected. I wonder how many other ideas that people tried before now work if we replace the linear layer with transformer layers.