Fifteen-week sequence
| 1 |
Linear Regression |
A tested linear regressor, a loss curve, and a short explanation of one optimization failure. |
| 2 |
Linear Classification and Shrinkage |
Binary and multiclass classifiers with a controlled L1/L2 comparison. |
| 3 |
Optimization and Regularization |
A controlled polynomial-regression comparison of learning rates, L2 regularization, and validation checkpoints. |
| 4 |
Feed-forward Networks and Backpropagation |
A gradient-checked MLP and an activation or width ablation. |
| 5 |
CNNs, ResNets, and Augmentation |
A small CNN and mini-ResNet evaluated under matched conditions. |
| 6 |
Transfer Learning, Fine-tuning, and U-Net |
A pretrained-encoder U-Net and a scratch-versus-transfer comparison. |
| 7 |
Diffusion Image Generation |
A tiny DDPM with reproducible samples and one schedule or sampling-step ablation. |
| 8 |
Deep Sets, Multiple-instance Learning, and GNNs |
An invariant set/MIL model and a small message-passing network. |
| 9 |
RNNs and GRUs |
A compact GRU sequence model and a long-dependency diagnostic. |
| 10 |
Transformer Core |
A reusable, unit-tested Transformer block used by Weeks 11–13. |
| 11 |
GPT, BERT, and Language Self-supervision |
A shared Transformer codebase with GPT and BERT configurations. |
| 12 |
Encoder–Decoder Models and Fine-tuning |
An encoder–decoder configuration and a fine-tuning comparison. |
| 13 |
Vision Transformers and Visual Self-supervision |
A mini-ViT plus a supervised or self-supervised representation comparison. |
| 14 |
Mamba, Selective State-space Models, and Distillation |
A simplified selective state-space model and a teacher–student experiment. |
| 15 |
Architecture Selection and Reproducible Comparison |
A final comparative project, reproducibility package, and individual architecture defense. |
Dependency chains
linear regression → classification → optimization/regularization → MLP/backprop
MLP → CNN/ResNet → U-Net → diffusion
MLP → Deep Sets/MIL → message-passing GNN
RNN/GRU → Transformer core → GPT/BERT → encoder–decoder → ViT
RNN/GRU + Transformer → selective state-space models → distillation
all model families → architecture selection and reproducible comparison