One running question: Module 03's risk flag barely beat the do-nothing baseline. Every keynote says deep learning finds the patterns the old models miss. Before the hospital pays for it: does it — on our patients? Spoiler, stated up front: probably not — this module is about finding out exactly why, and where the pitch is actually true. Arc 1 finds the ceiling a straight line can't cross (XOR) and the smallest thing that crosses it. Arc 2 shows you already built a neuron in Module 03, then traces a real 2-2-2 network forward by hand with a live interactive widget. Arc 3 derives backprop on a 2-parameter chain, checks it with PyTorch autograd, and reads a real loss landscape. Arc 4 builds the actual MLP on the diabetes data, runs straight into overfitting on purpose, then applies and compares three brakes. Arc 5 shows the same architecture family win clearly on MNIST pixels (unstructured data, deep learning's real strength) before closing on the honest four-model scoreboard back on the hospital data, and a multi-cause reading of AlexNet 2012.
Builds forward from Modules 01-03 — tidy data, statistics, and the patient-safe split are assumed. The diabetes anchor dataset, its 9-feature set, and Module 03's exact hand-fit logistic-regression weights are reused verbatim; MNIST is introduced fresh in Arc 5 as the module's one unstructured-data comparison.
This module's cells use torch (PyTorch, CPU) and the vendored MNIST files, which weren't in the
toolchain before. Re-run env/setup.sh once — it installs the newly pinned dependency, vendors
the MNIST files, and re-runs the smoke test.