Skip to content

Roadmap

Fourteen weeks, three blocks, three projects. Each block ends with something you built and can defend, and each one is less scaffolded than the last.

The sequence is deliberate. You spend the first four weeks barely modelling at all, because everything that quietly ruins a machine learning project happens before the model is trained.

Weeks 1–4 Data & the Pipeline
Weeks 5–9 Models & Optimization
Weeks 10–14 Representation & Systems
Block 1 Weeks 1–4

Data & the Pipeline

A model is the easy part.

01
Problem framing & baselines

Define a target, a metric, and a baseline before writing any model code.

02
Data quality & splitting

Choose the right split, and explore only the training data.

03
Feature engineering

Build preprocessing that is fitted on the training set alone.

04
Evaluation & error analysis

Pick a metric that matches the cost of error, and find where the model fails.

You can now

Frame a problem Split without leaking Engineer features Evaluate honestly

Project 1 · Raw data to a validated baseline

Block 2 Weeks 5–9

Models & Optimization

Understand what the fitting procedure is actually doing.

05
Linear models & the loss landscape

Derive linear and logistic regression from the loss up, and read the coefficients honestly.

06
Optimization

Implement gradient descent, and diagnose why it stalls or diverges.

07
Bias, variance & regularization

Read a learning curve, and reach for ridge or lasso for the right reason.

08
Tree-based models

Tune gradient boosting, and know when it beats a neural network.

09
Model selection & calibration

Run nested cross-validation, calibrate probabilities, and handle class imbalance.

You can now

Compare fairly Tune without cheating Read a learning curve Report variance

Project 2 · A defensible model comparison

Block 3 Weeks 10–14

Representation & Systems

Unlabeled data, and getting the thing out of the notebook.

10
Clustering

Cluster three ways and defend your choice of k without ground truth.

11
Dimensionality reduction

Run PCA via SVD, and read a t-SNE plot without being fooled by it.

12
Bridge to neural networks

Build an MLP as stacked linear models, and say honestly when it beats boosting.

13
ML systems & responsibility

Serve a model behind an API, and explain a single prediction to someone who is not you.

14
Synthesis & presentations

Present a complete system and defend its design under questioning.

You can now

Work without labels Reduce dimensions Ship a model Explain a prediction

Project 3 · Capstone

What this course does not cover

Deliberate omissions, so you know where to look instead.

Deep learning architectures. CNNs, RNNs, transformers. Covered in your Deep Learning course. Week 12 is the bridge to it.
Big data engineering. Spark, distributed training, warehousing. An adjacent discipline with its own course.
Reinforcement learning. Mentioned in Week 1 so you know where it sits on the map, not taught.

Before Week 1

Do this in the first days, not in Week 3

Everything you need is in the toolkit. Budget one afternoon.

Next: how the grade works, then Week 1.