Course / 02
Machine Learning
How a program learns a rule from examples and how to tell whether it will work on new data: from a line through points to gradient boosting and neural networks. Each note either fixes a limitation of the previous method or changes the task, and says which.
Before you start: the Probability Theory course, or at least expectation, conditional probability and maximum likelihood; vectors, matrix products and derivatives; reading short NumPy code. Every note ends with a Code section that runs as written. Learning Without Labels and Beyond Classification branch off the supervised route: they change the task rather than replace a classifier.
- 01
The Learning Problem
What it means to learn from examples, and how to know it worked.
After this section: Define the loss, the split and a baseline before fitting, and spot leakage in a pipeline.
4 of 4 notes published - 02
Linear Regression
A linear model fitted by formula and by descent, without fitting the noise.
After this section: Fit a linear model in closed form and by gradient descent, keeping the population target, the empirical loss, the optimizer and generalization apart.
4 of 4 notes published - 03
Linear Classification
Separating classes with a hyperplane, and measuring the errors that matter.
After this section: Tell a score from a probability and a calibrated probability from a threshold, and choose a metric that matches the cost of errors.
6 of 6 notes published - 04
Trees and Ensembles
Nonlinear models from simple rules, and why averaging and boosting them works.
After this section: Explain what a tree can express, why averaging trees lowers variance and how boosting fits what the ensemble still gets wrong.
4 of 4 notes published - 05
Neural Networks
Learning the features together with the model.
After this section: Trace one forward and backward pass by hand and diagnose a training curve that fails.
3 of 3 notes published - 06
Learning Without Labels
Structure in data when there is no target.
After this section: A change of task: find clusters, mixtures, low-dimensional structure and anomalies when there is no target.
4 of 4 notes published - 07
Beyond Classification
When the answer is an order or a recommendation.
After this section: Another change of task: learn an order or a recommendation, with metrics and splits that match how the list is used.
2 of 2 notes published