Labs

Each lab in this unit gives you hands-on experience with a different area of machine learning, working with real datasets and building systems that actually learn.

Estimation

Build your first models: start with an interactive toy where you fit a line to data by hand, discovering what "parameters," "loss," and "training" mean before those words are attached to any dataset. Then apply the same ideas—and the fit/predict pattern you'll use throughout this unit—to a large US public health survey, predicting income from health and lifestyle data using linear and multiple regression, and seeing firsthand how adding predictors can lead to overfitting. The dataset gives the technical work real stakes, setting up a serious discussion about bias, causation vs. correlation, and when it is appropriate to use personal data to build predictive models.

Classification: Features

Build a spam detector from scratch. Starting with hand-written rules, you will progress to designing numerical features and training a logistic regression model to weight them automatically, before stepping back to see how a Bag of Words model treats the entire vocabulary as features at once. Along the way you will learn how to evaluate a classifier honestly—using precision, recall, and F1 rather than raw accuracy.

Classification: Networks

Take on a harder problem: classifying handwritten digits from images. Hand-designed features, which worked well for spam, struggle here—and that failure motivates a new set of tools. You will hand raw pixel values to a model directly, then train multi-layer perceptrons and convolutional neural networks, learning why CNNs are well suited to spatial data and tuning hyperparameters to push past 95% accuracy.

Reinforcement Learning

Train a computer to play games without telling it the rules. You will implement Q-learning from scratch for a small grid game, then study a sequence of real training experiments on Snake to understand why what the agent sees—its observation design—matters as much as the learning algorithm itself. Finally, you will run your own training experiments on Frogger, changing hyperparameters and documenting what worked and why.