Intro to AI & ML, taught properly
Every concept from the course slides and notes — rebuilt from the ground up with intuition, analogies, full derivations, worked examples, interactive figures and guided problems.
The slides state results. They rarely teach them: a definition appears, a symbol is introduced without being named, a proof step says TODO, a poll asks a question and the answer lives in a clicker app you can't go back to.
This guide works the other way round. Every idea starts from something concrete, gets an analogy, then a definition with every symbol spelled out, then a numeric example computed all the way through, then the full derivation, then problems you solve yourself with the steps hidden until you want them. Nothing is left as an exercise to the reader.
Where to start
Scope is lectures 1–10 plus recitations 1–5 — no neural networks. Start with the scope checklist and two-day plan, then work the recitation problems, which are the staff's own.
Midterm 1 — scope & plan
What is and isn't on it, every topic mapped to a chapter and its problems, a two-day learning plan with a day for practice tests, and what deserves space on your one handwritten page.
Recitation problems, worked
Every problem from recitations 1–5 with full solutions in progressive steps — problem formulation, α-β traces, CSPs, vector derivatives, multiclass logistic regression.
How to use this guide
The dependency map between chapters, what the callouts mean, what to memorise versus re-derive, and which source documents each chapter came from.
Practice problem set
Exam-shaped problems across the whole midterm scope — search and A* through to regularization and cross-validation — each broken into steps you reveal one at a time.
Part I — Search
How an agent finds a plan: in a world that just sits there, in a world with an opponent trying to beat you, and in a world where the answer isn't a path at all.
Search problems, BFS, DFS & UCS
State spaces, search trees versus state-space graphs, the frontier, and the three uninformed strategies with their completeness and complexity worked out rather than tabulated.
Heuristics, Greedy search & A*
$f = g + h$, admissibility, consistency, the full A* optimality proof line by line, why graph search needs more than admissibility, and how to invent heuristics by relaxation.
Minimax, α-β pruning & Expectimax
Game trees, why the answer is a policy and not a plan, what the α-β cutoff test actually means, evaluation functions under a horizon, and chance nodes.
Constraint Satisfaction Problems
Identification instead of planning: backtracking search, forward checking, arc consistency and AC-3, and why MRV wants to fail fast while LCV wants to succeed fast.
Part II — Machine Learning
How an agent learns a function from examples: the framework, then two model families, then the optimization that makes either of them computable.
How an ML problem is set up
$\mathcal{X}$, $\mathcal{Y}$, $h$, loss, risk, empirical risk, ERM, generalization and the i.i.d. assumption — plus a notation table you can keep coming back to.
Decision trees
Stumps, majority vote, the recursive construction, greedy splitting, axis-aligned decision boundaries, and every extension the notes list.
Entropy & mutual information
Twenty questions, why $\log_2 \tfrac1p$ is the fundamental quantity, conditional entropy, information gain, and why it beats training error as a splitting criterion.
Linear regression
Fitting a line by completing the square, the normal equations $\theta^\star = (X^\top X)^{-1}X^\top y$ derived in full, and both corner cases answered.
Optimization: GD & SGD
Why the update rule looks the way it does, what the learning rate does to convergence, gradients of vector functions, local minima, and stochastic and mini-batch descent.
Feature engineering & overfitting
Linear in the parameters, not the input: polynomial and abstract transforms $\phi$, and the price you pay in overfitting when you add too many.
Logistic regression
From thresholding to the sigmoid, probabilities instead of hard decisions, $\sigma' = \sigma(1-\sigma)$, cross-entropy, the surprisingly clean gradient, convexity, and softmax.
Regularization & cross-validation
Why ERM always overfits, ridge vs LASSO (and why only one gives exact zeros), hyperparameters, holdout and $k$-fold cross-validation, and feature engineering vs feature learning.
Neural networks
Neurons and activations, a trainable three-neuron network, fully connected layers and parameter counting, backprop by hand, and the multivariate chain rule on computation graphs.
Neural networks II: backprop
Universal approximation, local minima, debugging under- vs overfitting, then backprop from a width-1 chain to the linear layer's $W^\top\boldsymbol\delta$ and $\boldsymbol\delta\mathbf a^\top$.
Computer vision: CNNs & ResNets
Why FC nets fail on images, convolution by hand and interactively, pooling and the CNN pipeline, AlexNet, and the vanishing-gradient puzzle that skip connections solve.
Maximum likelihood estimation
Likelihood vs probability, why we take logs, MLE for coins, dice and Gaussians (with the Lagrange step the notes skip), and why cross-entropy and squared error are negative log-likelihoods.
Formula cheat sheet
Every definition and formula in the course on one printable page, plus the ten traps most likely to cost you marks.
The shape of the course
Two halves that look unrelated and aren't. Search asks how do I find a good answer in a space too large to enumerate? Machine learning asks exactly the same question — the space is the set of hypotheses, and gradient descent is the search algorithm.
Built from your course documents
Everything here traces back to the professor's own notes and decks, so you can cross-check any claim against the original:
| Source document | Becomes |
|---|---|
07280_F26_Lec2_Informed_Search(-runover) | Chapters 1–2 |
07280_F26_Lec3_Adversarial_Search, 07280_S26_Notes_Adversarial_Search | Chapter 3 |
07280_F26_Lec4_CSPs | Chapter 4 |
07280_F26_ML_formulation, 07280_F26_Notes_ML_Prob_Decision_Trees §1 | Chapter 5 |
07280_F26_Lecture_Notes_Decision_Trees, …ML_Prob_Decision_Trees §2 | Chapters 6–7 |
07280_F26_Lecture_Notes_Linear_Regression, 07280_F26_Notes_Optimization_and_Linear_Regression | Chapter 8 |
07280_F26_Lecture_Notes_Optimization | Chapter 9 |
07280_F26_Notes_Feature_Eng_and_Logistic_Reg | Chapters 10–11 |
07280_F26_Lecture_Notes_Logistic_Regression | Chapter 11 §7–8 |
07280_F26_Lecture_Notes_Model_Selection | Chapter 12 |
07280_F26_Notes_Neural_Networks | Chapter 13 |
Several figures in the source PDFs are images that don't survive text extraction. Where a chapter needed one of those — the Iris comparison tables, the expectimax numbers, the overfitting panels — it has been reconstructed so the idea still lands, and the page says so at the point of use. Check those specific figures against your own copy of the slides before you rely on exact numbers.