Choose the right neural generative model for your problem
summary
Advice hinges on concrete tradeoffs: pick models by output structure and whether an objective exists. Continuous outputs favor direct nets or autoencoders for stable latent-space sampling. Photo‑real images use GANs — spectacular results but fragile to train. Discrete, highly‑constrained objects use hypergraph grammars or grammar‑masked decoding to guarantee validity; molecule work pairs grammars with reinforcement learning (PPO) and succeeds at valid candidates. AlphaZero style MCTS plus neural nets works but demands hundreds of CPUs/GPUs.
AI-written summary · about
chapters (33)
- 0:00Generative models intro
- 1:09Avoiding generative models
- 2:13Model selection questions
- 3:40Output space structure
- 5:05Optimization patterns
- 6:08Encoder-decoder fixed latent
- 8:18Encoder-decoder variable seq
- 9:46Generative adversarial networks explained
- 11:12Reinforcement learning for generation
- 12:57Problem structure mapping
- 14:02Discrete objectives handling
- 15:04Molecule optimization latent
- 16:53RL in discrete spaces
- 17:33Generation without objectives
- 19:21Encoder-decoder architecture
- 21:54Transformers for inputs
- 23:18Decoder for continuous outputs
- 24:01Beam search decoding
- 24:23Output transformation
- 26:13Grammars for constraints
- 27:19Software as output transform
- 28:23Practical examples closing
- 28:45Q&A molecular optimization
- 31:06Search strategies
- 32:36Image generation classes
- 33:18Semi-supervised limited data
- 35:28Grammar-based RL
- 36:29Training with grammar rules
- 38:30RL black-box rewards
- 40:22Multi-reward prediction
- 41:28Reward signal encoding
- 42:03Stabilizing GANs
- 43:31Closing Q&A
related talks
43mHitchhiker’s Guide to Generative Adversarial Networks (GANs)
53mGenerative AI: Overview and use cases
47mTrapping the Accident
40mOverview of Generative Adversarial Networks (GANs)
01h 00mTutorial on Deep Learning IV
37mHarnessing the power of Generative Adversarial Networks (GANS) for supervised learning
52mDeep Generative Models And Unsupervised Methods For Inverse Problems
30mVisual data abundant relevant labelled cheap Pick two