📗 -> 04/30/25: ECS189G-L14

GAN Slides
Computer Vision Slides

🎤 Vocab

❗ Unit and Larger Context

Continuing with GAN

A quick summary of GAN

Discriminative models vs generative models

Examples of discriminative models
Can we just use auto-encoders as the generative model?
Why we need to study generative models?

Generative Adversarial Model

The game between generator and discriminator
How to train the model iteratively

What can GAN do?

Some examples
More to be explored by you!
Source code tutorials: https://pytorch.org/tutorials/beginner/dcgan_faces_tutorial.html

✒️ -> Scratch Notes

REVIEW GAN LOSS FUNCTION, FINAL QUESTION ON IT

Begin class by reviewing how to define and fix D and G loss functions in GAN

Iterate by fixing G, then optimizing D
then fixing D, then optimizing G

Do this iteratively to efficiently optimize them both to convergence

Stages:

Poorly fit G and D
G will approach the underlying data distribution
D will adjust making some strategy to differentiate the two, even if its mostly probabilistics
G will fit, be indistinguishable from the data
D will start guessing randomly, will be unable to tell the data apart

GAN Uses

Turn input into generated images, not just feeding noise:

Sketch bag to real bag
Sketch face to real face
Face to emoji
Input face, aging
Text to image

Moving on from GAN

Computer Vision

The science and technology that make machines “see”
Image data can be diverse
- B/W, gray scale, RBG
- Videos (sequences of images)
  - Normally at least 24 FPS
- Depth images, like Lidar with point cloud
- High-dimension medical scanning images
  - A high dimensional image tensor (multi-bytes per pixel)

Vision for humans

Very powerful vision system inherited from our ancestors

Evolving for millions of years
Our brain is “initialized” to recognize items at our birth
We can perceive items in the 3D space effectively

Back to CV

Is computer vision that hard?

Yes…
Very hard for current computers

Need to make machine vision system from scratch

Choose how to see, camera vs. lidar
How to identify humans vs cars
How to learn, DL or other methods

But have advantages too: human vision weak to optical illusions (visual processing relative, contextual)

CV Problems

Image classification
Object
- Localization
- Detection
- Segmentation
Image
- Style transfer
- Reconstruction
- Super-resolution
- Synthesis
others…

🧪 -> 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?)

Vault

Explorer

ECS189G-L14