Lectures

Course logistics, a brief history of AI, and the PEAS model for understanding rational agents, which will help us conceptually organize the various problems we will tackle in the class.

Book chapters: Ch. 1 (Introduction), Ch. 2 (Intelligent Agents)

Note: The book covers much more than we are able to see in class. Unless otherwise specified, the slides should be self-contained for you to follow the course, and the book is a resource for going deeper into material that interests you.

Recap of the PEAS model and definition of search problems as a kind of static, deterministic, discrete environment; Uninformed Search methods (BFS, DFS, ID-DFS, UCS), Informed Search and heuristics (A* and variants), admissible heuristics and optimality, deriving heuristics from relaxations.

Live coding: [Github]

Book chapters: Ch. 3 (Solving Problems via Search)

Optional additional readings: Two papers using learned search heuristics: DeepCubeA (Rubik's Cube), BFS-Prover (theorem proving in Lean 4)

Definition of a CSP (Constraint Satisfaction Problem). Sudoku as a CSP. Our first factored state representation: states are now variable assignments. Backtracking search and heuristics for variable and value ordering (LCV, MRV), and inference (AC-3, forward checking, MAC).

Book chapters: Ch. 6 (Constraint Satisfaction Problems)

Optional additional reading: Training a Transformer to solve CSPs via local search: ConsFormer (ICML 2025)

Knowledge-Based Agents. Propositional Logic syntax and semantics. Truth-table method for deciding entailment. Satisfiability and validity. Inference rules, soundness. Proofs, completeness, refutation-completeness and resolution.

Book chapters: Ch. 7 (Logical Agents)

Optional additional reading: U-M Prof. Karem Sakallah and former PhD student João Marques-Silva just won an award from the SAT association for 30 years of their groundbreaking work on SAT solving: CSE article

Recap of propositional logic. The DPLL algorithm. First-order logic: syntax and semantics.

Book chapters: Ch. 8 (First-Order Logic)

Optional additional reading: A tutorial on Z3, the solver we mentioned in lecture.

Quantifying uncertainty with proability. Probability axioms. Inference and estimation. Bayes' rule, Bayesian learning and MAP inference. Maximum Likelihood Estimation. Independence and mutual independence. Naïve Bayes classifier.

Book chapters: Ch. 12 (Quantifying Uncertainty)

Naïve Bayes classifier. Bayesian Networks. Exact Inference: inference by enumeration, Variable Elimination. Approximate Inference: Rejection Sampling, Gibbs Sampling.

Book chapters: Ch. 13 (Probabilistic Reasoning)

Gibbs Sampling. Introduction to machine learning: hypothesis class, loss function, optimization. Linear regression case. Gradient Descent and SGD.

Book chapters: Ch. 19 (Learning from Examples), in particular 19.1, 19.2, 19.4, 19.6

Classification. Estimation and Approximation error. PAC Learning framework. Non-linear features. Neural Networks.

Book chapters: Ch. 19 (Learning from Examples)

Backpropagation. Automatic differentiation. Optimizers: SGD + Momentum, RMSProp, Adam, AdamW. Learning rate scheduling.

Book chapters: Ch. 21 (Deep Learning). We covered 21.1, 21.2, 21.4.1. Material on optimizers not covered in the book.

Learning rate scheduling, initialization, normalization, residual connections. Convolutional neural networks.

Book chapters: Ch. 21 (Deep Learning).

Training neural networks on natural language. text tokenization, and byte-pair encoding, and embedding layers. Recurrent Neural Networks. Attention. Multi-head self-attention and the Transformer architecture. Positional encoding.

Book chapters: Ch. 24 (Deep Learning for Natural Language Processing).

Transformer architecture. GPT models (GPT, GPT-2, GPT-3) and new capabilities gained from scaling (transfer, zero-shot, few-shot learning). Scaling laws for Large Language Models. Systems challenges in large-scale training.

Review quiz on Gradescope.