Скачать или смотреть Session 3 Part 2 - Deep Generative Modeling

Welcome to Session 3.2 – Deep Generative Modeling with Autoencoders & GANs

In this lecture and lab walkthrough, we explore deep generative models, including autoencoders, variational autoencoders (VAEs), and generative adversarial networks (GANs), as part of the How to Build and Train Your AI Model workshop series. We’ll cover how these models can learn from unlabeled data to generate realistic images and representations, with real-world applications in facial generation, bias reduction, and creative AI.

What You’ll Learn in This Video:
Understanding the difference between supervised and unsupervised learning
What latent variables are and how they capture meaningful representations
How autoencoders and variational autoencoders work
Reparameterization trick and regularization in VAEs
Introduction to GANs and the adversarial training loop
Real-world applications of GANs including CycleGANs and Pix2Pix
Generating new images, translating domains, and detecting bias

Hands-on Learning:
Visual intuition behind encoding/decoding with VAEs
Real-time comparison of generated vs. real faces
Understanding and tweaking the loss function and latent space
GAN training flow and adversarial loss
Combining AI-generated elements for novel outputs (image interpolation)

Resources & Links:
Course and lab materials: https://www.InspiredK.org
GitHub repository: https://github.com/InspiredK-organiza...

YouTube Chapters & Timestamps:
0:01 – Introduction to Deep Generative Modeling
0:18 – Can You Tell Which Face is Real? (None Are!)
1:00 – Supervised vs. Unsupervised Learning
2:05 – Density Estimation & Sample Generation
3:00 – Real-World Importance of Generative Models
4:28 – Outlier Detection & Bias Mitigation
5:09 – Autoencoders & Latent Variables
6:13 – Plato’s Cave & Latent Representations
7:02 – Encoder & Decoder Architecture
8:04 – Reconstruction Loss (No Labels Needed)
9:13 – Latent Space Dimensionality & Quality
10:04 – Introducing Variational Autoencoders (VAEs)
11:00 – Sampling from Mean & Standard Deviation
12:02 – The Role of Regularization & the Gaussian Prior
13:13 – KL Divergence & Loss Functions in VAEs
14:30 – Visualizing Latent Space: Complete vs. Unregularized
16:23 – Laten Space Interpolation & Feature Mapping
17:02 – The Reparameterization Trick
18:35 – Perturbation, Disentanglement, and Diversity
20:11 – Summary of Variational Autoencoders
21:19 – Introduction to Generative Adversarial Networks (GANs)
22:30 – Generator vs. Discriminator: The Cat & Mouse Game
23:44 – Concrete Training Example: GAN in Action
25:13 – When the Discriminator Can No Longer Tell the Difference
26:02 – Training Loss for G & D in GANs
27:14 – Latent Space Interpolation: Duck + Finch = ?
28:21 – Real Applications: Deepfakes, Maps, Satellite Data
29:28 – Conditional GANs & Pix2Pix Mapping
30:54 – CycleGANs: Mapping Between Two Domains
31:43 – CycleGAN in Action: Horse → Zebra
32:01 – Audio Translation, Music, and Spectrograms
32:59 – Summary: VAEs vs. GANs & Creative Generation

🎨 Try This: Create your own face using a GAN lab
💬 Comment Below: What would you generate with an AI?
🔔 Subscribe & Turn on Notifications for More Labs & Lectures