This is my study hub for the MPL master’s course. Each lecture gets its own page with explanations, key concepts, math, and examples I worked through to actually understand the material — not just memorize it.
I use the SlideLink tool I built to automatically align these notes with the lecture slides.
How to Use These Notes
- Reading linearly works well — each lecture builds on the previous.
- Each page has a mental model section (big picture first), then details.
- Look for the Example blocks to build intuition.
- Math is included where needed, but always paired with plain-English explanations.
Lectures
| Topic | Core Idea | Example Model / Use Case |
|---|---|---|
| (y-01) Introduction | What is ML? Loss, optimization, training loops | Logistic Regression (Iris dataset) |
| (y-02) Convolutional Neural Networks | Spatial feature extraction with filters | Simple CNN (MNIST digit recognition) |
| (y-03) Vision CNNs | AlexNet, VGG, ResNet, EfficientNet | ResNet-50 (ImageNet classification) |
| (y-04) Recurrent Neural Networks | Sequences, LSTMs, GRUs, vanishing gradients | LSTM (Sentiment analysis / Stocks) |
| (y-05) Transformers | Attention is all you need | BERT / GPT (Machine Translation) |
| (y-06) Vision Transformer (ViT) | Patches + Transformers = vision | ViT-Base (Large-scale visual recognition) |
| (y-07) Multimodal Learning | CLIP, image+text, cross-modal alignment | CLIP (Zero-shot image classification) |
| (y-08) Interactive Machine Learning | Humans in the loop | Active Learning / GNN (Sudoku solver) |
| (y-09) Generative AI & VAE | Latent spaces and variational inference | VAE (Face generation / Reconstruction) |
| (y-10) GANs | Generator vs. Discriminator | StyleGAN (Synthetic high-res faces) |
| (y-11) Reinforcement Learning | Rewards, policies, Q-learning | Q-Learning / PPO (Game playing / Atari) |
| (y-12) Diffusion Models | Denoising as generation | Stable Diffusion (Text-to-image generation) |
| (y-13) Explainable AI (XAI) | Why did the model decide that? | Grad-CAM / SHAP (Debugging model bias) |
Applied MLP: Case Studies
Practical applications of theory to real-world technical problems.
(y-) Case Study: The Sudoku GNN
Core Idea: Most neural networks work on Euclidean data (grids of pixels or sequences of text). Sudoku is better represented as a Graph, where the rules of the game define the edges (connections) between cells.
Mental Model: Message Passing as Constraint Propagation
- In traditional Sudoku solvers, you look at a cell and “propagate” the constraints from its row, column, and box to eliminate possibilities.
- In a GNN, this is exactly what Message Passing does. Each node (cell) sends its current “state” (clues and predictions) to its neighbors. After a few rounds of updates, each node has “seen” enough of the board to make a classification.
💡 Intuition: Why use a Graph instead of a CNN?
- A CNN is limited by its receptive field; it can only “see” a small 3x3 or 5x5 area at a time.
- A GNN with explicit edges for rows and columns has a receptive field of 1 for all constraints. This “shortcut” for relational reasoning makes the learning problem significantly easier.
Quick Reference: Key Concepts
| Concept | Where It Appears |
|---|---|
| Backpropagation | L01, L02 |
| Attention | L05, L06, L07 |
| Latent Space | L09, L10, L12 |
| Sequential Data | L04, L05 |
| Generative Models | L09, L10, L12 |