Machine Learning

study guide

How to Work Through the Book

Do not read this course as trivia about AI. Read it as a sequence of mental models, mathematical objects, and experiments. The goal is not to remember every term. The goal is to see a new machine-learning system and know what questions to ask.

Three Reading Tracks

First-year orientation

For readers who want the big picture before the math gets dense.

  1. Read the hook, idea, demo, takeaways, and coda first.
  2. Skip advanced GoDeeper boxes on the first pass.
  3. Return to the math details after you can explain the demo in your own words.

Graduate core

For readers who want the machinery behind the intuition.

  1. Work every ProblemSet without looking back at the prose.
  2. Run the browser labs and modify at least one assumption.
  3. For each phase, write the objective, assumptions, estimator, and failure mode.

Builder track

For readers who want to turn the course into working systems.

  1. Treat every demo as a specification for a small implementation.
  2. Replace toy data with a real dataset when the chapter ends.
  3. Use the projects page as a portfolio spine.

Phase Checks

At the end of each phase, pause and make the promise falsifiable. If you cannot do the check without rereading the chapter, the concept is still vocabulary, not skill.

  1. Phase 1 The Learning Loop

    You can describe any supervised learning system using data, model, loss, optimizer, and generalization.

  2. Phase 2 Generalization and Measurement

    You can design train/validation/test splits, choose metrics, and explain why held-out performance matters.

  3. Phase 3 Linear Models and Optimization

    You can derive how a line becomes a classifier and how a local update changes the decision boundary.

  4. Phase 4 Deep Networks

    You can explain how a scalar loss teaches many layers and why training stability is an engineering problem.

  5. Phase 5 Representations Across Space and Time

    You can recognize the inductive bias behind CNNs, RNNs, and attention instead of memorizing architectures.

  6. Phase 6 Foundation Models

    You can trace how next-token training becomes a general interface for language, images, retrieval, and tools.

  7. Phase 7 Agents, Alignment, and Evaluation

    You can reason about policies, preference models, eval suites, reward hacking, and agent reliability.

Notation Ledger

Machine learning notation is overloaded across papers. This course uses the following ledger unless a chapter says otherwise.

Symbol Meaning Used in
x One input example or feature vector. All supervised learning chapters.
y The target label, value, token, reward, or desired output. Losses, metrics, generation.
f(x; theta) A model with parameters theta applied to input x. Prediction, optimization, neural networks.
theta The parameters that training changes. Gradient descent and model fitting.
L(y_hat, y) A loss for one prediction compared with one target. Learning loop and optimization.
R(f) Expected risk, the average loss under the real data distribution. Generalization.
R_hat(f) Empirical risk, the average loss on the sample we have. Generalization.
grad_theta L The direction each parameter should move to change the loss. Gradient descent and backpropagation.
W, b Weights and bias in a linear layer or neuron. Linear models and neural networks.
h Hidden state or hidden representation. Deep networks, RNNs, transformers.
q, k, v Query, key, and value vectors used by attention. Attention and transformers.
pi(a | s) A policy giving action probabilities in a state. Reinforcement learning and agents.

Comprehensive Exam Prompts

These prompts are the course's exit criteria. A strong answer should use concrete examples, the right metric, and at least one failure mode.

  1. Given a new ML problem, write the data schema, model family, loss, metric, and evaluation split.
  2. Explain why a lower training loss can be bad news.
  3. Trace one prediction through a neural network forward pass and one error through the backward pass.
  4. Compare CNNs, RNNs, and attention by the structure each assumes about the input.
  5. Design a retrieval or tool-using AI system and name where it can fail.
  6. Write an eval plan that would catch both capability regressions and unsafe behavior.