Large Scale Representation Learning In-the-wild
summary
Claims that unsupervised and self‑supervised methods match or beat supervised baselines on ImageNet and cut label needs by orders of magnitude, showing an example of ImageNet 75% vs 45% when labels are scarce. Describes concrete tools: autoencoders with a 30‑dim bottleneck, GANs and BiGANs that generate believable faces and pizzas (with visible artifacts), and contrastive learning and transformer generative models (ICLR 2022) that produce class‑conditioned samples. Notes web‑scale unlabeled corpora and practical augmentation recipes used for training.
AI-written summary · about
chapters (10)
- 0:00Representation learning intro
- 4:01Supervised representation learning
- 6:30Why unsupervised learning
- 9:22Self-supervised learning
- 13:40Autoencoders compress data
- 21:09Generative Adversarial Networks and representation
- 28:18Beyond GANs and contrastive
- 33:01Contrastive learning
- 38:21Advanced models future
- 39:52Q&A
related talks
40mVisual concept learning from few images
30mVisual data abundant relevant labelled cheap Pick two
36mSemi-supervised deep learning with GANs
04mUsing Computer Vision to Understand Big Visual Data
01h 02mTutorial on Deep Learning I
28mChallenges & future trends of Deep Learning
41mDeep Learning to Solve Challenging Problems
12mENHANCE! Upscaling Images with Neural Networks