06 · Machine learning · 2025 — 2026
ML from scratch — Stanford CS229
The problem sets of Stanford's CS229 in pure NumPy — from logistic regression and kernel SVMs to a neural network with manual backpropagation, EM and reinforcement learning. Every algorithm re-derived from the math.
My solutions to the problem sets of Stanford's CS229 (Machine Learning), in pure NumPy with no ML framework. The rule: re-derive every algorithm from the mathematics before ever calling a library that does it for you.
Implemented: logistic regression with Newton's method, Gaussian discriminant analysis, Poisson and locally weighted regression; the perceptron, a kernelized SVM and a naive Bayes spam filter; a neural network with manual backpropagation; Gaussian mixture models trained with EM; independent component analysis; and reinforcement learning on the cart-pole problem.