BSDA5002 — Generative AI Foundations
231 words
1 min read
Visual companion
Python
Type and operator map
Python Week 1: the first filter for runtime behavior
View
Revision summary
What this note is really saying
Short form
# BSDA5002 — Generative AI Foundations > **Course**: Degree Elective (4 credits) **Topics**: GANs, VAEs, Diffusion Models, DDIMs, Autoregressive Models, Evaluation **Files**: 10 comprehensive topic files ## Course Index File Topic Key Concepts 01-generative-model-overview Generative Models Overview Taxonomy, likelih...
BSDA5002 — Generative AI Foundations
Course: Degree Elective (4 credits) Topics: GANs, VAEs, Diffusion Models, DDIMs, Autoregressive Models, Evaluation Files: 10 comprehensive topic files
Course Index
| File | Topic | Key Concepts |
|---|---|---|
| 01-generative-model-overview | Generative Models Overview | Taxonomy, likelihood-based vs implicit, MLE, i.i.d. assumption |
| 02-gans | Generative Adversarial Networks | Min-max game, Nash equilibrium, DCGAN, WGAN, mode collapse |
| 03-vaes | Variational Autoencoders | ELBO, reparameterization trick, KL divergence, Beta-VAE |
| 04-diffusion-models | Diffusion Models (DDPM) | Forward diffusion, reverse denoising, noise scheduling, U-Net |
| 05-ddim | Denoising Diffusion Implicit Models | Accelerated sampling, non-Markovian, deterministic inversion |
| 06-autoregressive | Autoregressive Models | PixelCNN, PixelRNN, causal masking, sequential generation |
| 07-conditional-generation | Conditional Generation | Class-conditioning, guided diffusion, classifier-free guidance |
| 08-evaluation-metrics | Evaluation Metrics | FID, IS, likelihood evaluation, precision-recall, diversity |
| 09-information-theory | Information Theory | Entropy, KL divergence, mutual information, cross-entropy |
| 10-math-foundations | Math Foundations | SVD, eigendecomposition, matrix calculus, ELBO derivation |
Key Formulas
| Formula | Description |
|---|---|
| GAN objective | |
| $\mathcal{L}{VAE} = \mathbb{E}{q_\phi(z | x)}[\log p_\theta(x |
| DDPM forward process | |
| $\mathcal{L}{DDPM} = \mathbb{E}{t, x_0, \epsilon}[\ | \epsilon - \epsilon_\theta(x_t, t)\ |
| $FID = \ | \mu_r - \mu_g\ |
| $KL(P\ | Q) = \sum_x P(x) \log\frac{P(x)}{Q(x)}$ |
Exam Weightage
| Topic | Quiz 1 | Quiz 2 | End Term |
|---|---|---|---|
| GANs | ★★★★★ | ★★★ | ★★★★ |
| VAEs | ★★★★ | ★★★ | ★★★★ |
| Diffusion Models | ★ | ★★★★★ | ★★★★★ |
| DDIMs | ★ | ★★★★ | ★★★ |
| Autoregressive Models | ★ | ★★★★ | ★★★ |
| Evaluation Metrics | ★ | ★★★★ | ★★★★ |
| Math Foundations | ★★★ | ★★★★ | ★★★★★ |
Cross-Course Links
- BSDA5004 (LLMs): Autoregressive generation, attention mechanisms
- BSDA5006 (DL-CV): GANs for image generation, U-Net architecture
- BSMA1003 (Maths 2): Linear algebra foundations
- BSMA1004 (Stats 2): Probability distributions, MLE Join Discord PreviousMath Foundations