Visual data abundant relevant labelled cheap Pick two
summary
Describes concrete tools and limits: use transfer learning, classic augmentations and use adversarial augmentations in both training and testing to probe robustness; train GANs/VAEs but expect they require a lot of seed images and produce alignment and texture artifacts; StyleGAN produces realistic faces but yields occasional impossible teeth or glasses glitches; simulate full systems with CARLA or AirSim for reinforcement or sensor stacks but watch the domain gap; apply generative methods for anonymization rather than as a drop-in replacement for labeled real data.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 3:14Assumptions
- 4:58Augmentations for training
- 8:29Adversarial augmentations
- 13:45Anonymization and transformations
- 14:50Generative models
- 20:49StyleGAN limitations
- 23:16Feature disentanglement
- 25:46Deepfakes for privacy
- 26:07Simulation and virtual data
- 29:24Recap and approach
related talks
41mLarge Scale Representation Learning In-the-wild
13mNavigating Generative AI and Its Risks
40mOverview of Generative Adversarial Networks (GANs)
43mHitchhiker’s Guide to Generative Adversarial Networks (GANs)
40mVisual concept learning from few images
01h 18mComputer Vision after the Victory of Data
01h 17mDeep generative models for image...
28mChallenges & future trends of Deep Learning