ECS189G-L13

📗 -> 04/28/25: ECS189G-L13

---

GAN Slides

🎤 Vocab

❗ Unit and Larger Context

Small summary

✒️ -> Scratch Notes

Discriminative models vs Generative models

Discriminative models

Based on seen data (training data)
Learn a mapping to fit the training data set

• Assumption: distribution of training data is similar to a underlying distribution that governs all data, including unseen ones
  A MNIST classifier is an example of a discriminative model

Generative Models

We create a training dataset out of all data, and build a model that can generate a data instance sample. No ground truth.

AIGC - AI Generated Content
SAM - Segment Anything Model (META AI)

• segment-anything.com | includes an online demo
  Genesis - Combine generative models with physics engine
• genesis-embodied-ai.github.io

Many problems require generative models:

• Speech synthesis: text -> speech audio
• Machine translation: EN -> FR, CN -> JP
• Image segmentation: input image -> segmented regions

Auto-encoders

The second half (the decoding) of an auto-encoder is just like a generative model

Problems

• Doesn’t try to simulate real data, only copy inputs
  • If there are no inputs, cannot generate embeddings
• Cannot generalize well to unseen images
  • Minor change to inputs can generate very different outputs
• Limited application in diverse learning settings
  • AE are unsupervised, difficult to generalize

Games

Game theory is the stud of mathematical models for strategic interactions among rational agents

• Assumes: all agents are selfish and aim to maximize payouts

Nash Equilibrium - Each agent’s strategy is the ‘best’ possible for their opponent’s strategy

Prisoner’s Dilemma:

• For each of them, it is payoff maximizing in 1 match to betray.
  Zero-sum game
• No wealth is created or destroyed, any person’s gain is another’s loss
  Min-Max game

GAN - Generative Adversarial Network

A game between discriminative model and generative model

• Discriminator D: Distinguish data instances samples from data distribution from those generated by the generator G
• Generator G: Tries to trick the discriminator D by generating data instances that are as hard as possible to distinguish
  • Hard to distinguish Real data
• A min-max between D and G
  After running the model, you will get loss functions D and G

Objective Function:

• : Probability of x to be really D
• : Generation output by G on noise
  • Leads to
• The expectation:

🧪 -> Refresh the Info

Did you generally find the overall content understandable or compelling or relevant or not, and why, or which aspects of the reading were most novel or challenging for you and which aspects were most familiar or straightforward?)

Did a specific aspect of the reading raise questions for you or relate to other ideas and findings you’ve encountered, or are there other related issues you wish had been covered?)

🔗 -> Links

Resources

• Put useful links here

Connections

• Link all related words