Semi-supervised deep learning with GANs
summary
Presents a semi-supervised learning method using GANs in PyTorch to turn 100 labeled images plus ~13,000 unlabeled ones into a practical classifier. Uses a ResNet18 baseline (about 60% with 100 labels) and a GAN approach that trains a generator to produce complementary (intentionally poor) samples in feature space; this yields ~72% accuracy at 13 epochs and ~83% at 40 epochs (>20% improvement). Also reports that curated selection of labeled examples gives ~20% boost versus ~10% for random picks, and the training uses a modified loss instead of adding an extra output logit.
AI-written summary · about
chapters (14)
- 0:00Intro
- 2:40Welcome
- 2:50Outline
- 3:30Semi-supervised machine learning
- 5:08Why semi-supervised learning
- 6:44How it works
- 9:15GAN architecture
- 15:37How good are GANs
- 16:48Basic idea
- 19:47Three types of data
- 21:20Bad GANs
- 23:47Practical
- 28:20Morals
- 32:25Questions
related talks
01h 27mSemi-supervised learning with GANs
37mHarnessing the power of Generative Adversarial Networks (GANS) for supervised learning
41mLarge Scale Representation Learning In-the-wild
40mVisual concept learning from few images
43mHitchhiker’s Guide to Generative Adversarial Networks (GANs)
01h 00mTutorial on Deep Learning IV
40mOverview of Generative Adversarial Networks (GANs)
30mVisual data abundant relevant labelled cheap Pick two