Harnessing the power of Generative Adversarial Networks (GANS) for supervised learning
summary
Builds a face-frontalization pipeline: an encoder-decoder generator with a 512-dimensional bottleneck and ~5M trainable parameters, trained on about 700,000 image pairs using PyTorch and NVIDIA tooling. Training uses GPU preprocessing with DALI and a Tesla P100, giving ~18 minutes per epoch versus 8.5 hours on 4 CPUs. Adding a discriminator (GAN) during training sharpens fine details—eyelashes, eyelids, nose corners—so models generalize better at inference; code and experiments live on the Scaleway GitHub.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 0:14Agenda
- 1:26Generating content with AI
- 2:37Supervised vs unsupervised learning
- 9:07Learning pipeline
- 13:31CPU vs GPU
- 20:54Generator Architecture
- 26:45Training
- 27:31Results
- 29:14Inference
- 30:45Change Return
- 32:47Visual Quality
- 34:06Gun Example
related talks
40mOverview of Generative Adversarial Networks (GANs)
43mHitchhiker’s Guide to Generative Adversarial Networks (GANs)
24mA Journey through Generative AI's Frontier Python Libraries
47mUnmasking Generative Adversarial Networks
01h 27mMaking Faces: Conditional Generation of Faces using GANs...
36mSemi-supervised deep learning with GANs
12mENHANCE! Upscaling Images with Neural Networks
48mUnveiling the secrets of a machine learning model