A set of interactive lessons built around one lecture deck — the pipeline that turns a dataset into a deployed decision.
| # | Lesson | Covers |
|---|
| 1 | From data to decisions | The core pipeline: D→A→f^_D, then x→f^_D→s→δ→a |
| 2 | Beyond the straight line | Linear vs. softmax regression vs. kNN — parametric vs. nonparametric |
| 3 | Evaluation: how good is "good"? | MSE, ROC/AUC, confusion matrices, thresholds, and group fairness |
| 4 | Build the policy | Interactive review exercise — the whole arc, one decision at a time |
| 5 | Shape and stacking | Perceptron, linear separability, XOR, and the multi-layer perceptron |
| 6 | How the stack learns | The perceptron's simple nudge rule, credit assignment, backpropagation, and why activations went smooth |
| 7 | The hierarchy, and where transformers fit | AI ⊃ ML ⊃ neural network ⊃ transformer, attention, GPT, and where models still break |
| 8 | How much data is "huge"? | Grounding "trained on a huge pile of text" in real, sourced numbers |
| 9 | The scale surprise | Same training objective, huge scale — why capability like coding and translation shows up unasked |
| 10 | Historical context | The people, dates, and accidents behind lessons 0005–0009, in order |
| # | Lesson | Covers |
|---|
| 11 | Homework 1: the map | How each problem maps to lessons 0001–0010, plus the new notation at a glance |
| 12 | Thresholds and policy | Confusion matrix, TPR/FPR, capacity-constrained thresholds, grouped validation splits |
| 13 | Log loss and one gradient descent step | Sigmoid, log loss, the gradient, one GD update, and why L2 regularization is needed under separation |
| 14 | Softmax and deployment | Categorical labels, softmax probabilities, log loss vs. classification error, a confidence threshold that routes to a human |
| 15 | Regularization can prefer a shortcut | L1 vs. L2 penalty arithmetic and geometry, spurious features and audits |
| 16 | The V-shaped gate | Bonus — two ReLU units composing into |x| |