Handouts
Full class walkthroughs as single-page documents. Each one renders every figure from real training runs, ships the code to reproduce it, and includes solved checkpoints.
They complement the class pages rather than replace them: each goes deep on one topic and works both as a companion during class and as a reference afterwards.
Written in Portuguese
The handouts are course material for the in-person classes and are written in Brazilian Portuguese. The code, the library names and the figures are language-independent.
-
Regularization in practice: making a network stop memorizing
digits — 1 797 handwritten digits at 8 × 8 pixels, of which only 300 are used for training, and 30 of those carry a deliberately wrong label.
One architecture, 64 → 128 → 128 → 10, held fixed while five techniques are switched on one at a time and measured: early stopping with restore, L2 across seven values of λ, dropout across six values of p, ×5 augmentation, and label smoothing — with what each one costs when pushed too far. The weight norm tracked against the generalization gap, the confidence histogram and the reliability curve that accuracy hides, and a final stack that turns 88.8% into 96.4% on test without adding a single parameter. Written in NumPy, so no technique is a black box called by name.
~2 h · 8 figures · 17 measured runs · 8 checkpoints with solutions
Classroom use
Each handout has an anchored sidebar index, a copy button on every code block, and its own print stylesheet — the menu, buttons and interactive elements are dropped when printing or exporting to PDF.
The nine simulators for this chapter live on the class page — including one where you stack these same techniques and watch the decision boundary change in real time.