📗 -> 04/23/25: ECS189G-L11
sec_6_auto_encoder.pdf - Google Drive
sec_7_convolutional_neural_net.pdf - Google Drive
🎤 Vocab
❗ Unit and Larger Context
Summary of Section 6 - Auto Encoders
Auto-encoder
- What is auto-encoder?
- What is the model architecture?
- How to train the model? e.g., loss function definitions…
- What can auto-encoders do and what are their advantages?
- What are the potential applications of auto-encoders?
Variational Auto-encoder (VAE) - It is more like a probabilistic graphical model
- Why it is also called “auto-encoder”?
- What is VAE model architecture and learning process?
- (Optional) VAE learning with gradient descent via re- parameterization.
CNN
How to represent
Conventional Image Processing Operators
Convolution Operator
CNN Model Architecture & Pooling Layer
CNN Model Variants
ResNet: Deep CNN with Residual Learning
✒️ -> Scratch Notes
Variational Auto-Enconders (VAE)
Less similar to the previous auto-encoder variants
Based on: Bayesian ad Graphical Model
- Prior likelihood
- Likehood
- Posterior Likelihood:
x is the image
z is the embedding of the image
How do we generate a data instance
- Sample a
from a prior distribution - Generate a value
from a conditional distribution
Distribution parameter learning
Review
what?
Common Arches:
Img -> VAE encoder -> UNet -> VAE Decoder -> img
Text -> CLIP ? -> /\
CNN
Image representation:
- Black/White Image Matrix (1 or 0 matrix)
- Gray image Matrix (1 byte per pixel)
- Color Image Tensor (3 matrices, 1 byte for Red/Green/Blue)
Weakness of MLPs for 3 color images:
- Insufficient for CV taks
- Too many parameters to learn
- Input layer size for image of size m*n would be mn (gray images) or 3mn (color images)
- With more hidden layers to learn, even more to consider
- Discards spatial information, all into one vector of independent elements
- Assumes everything is iid (independent and identically distributed)
Common Image Processing Operators
-
Gaussian Blur
-
Sobel Operator (edge detection, look up sobel matrices)
-
Canny Edge Detector (edge detection thinning edges)
-
Image Blur
- Define a kernel, and apply to image A:
Gaussian Blur:
Sobel Operator:
Convolution
Convolution Mathematical Definition
An operator defined for two functions f(t) and g(t)
🧪 -> 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