Course Description:
This course introduces modern machine learning techniques for developing and deploying generative models across a range of data modalities. Topics include autoregressive models for text and image generation, such as transformers and GPT-style architectures, with sampling carried out sequentially from conditional distributions. The course also covers image generation approaches including variational autoencoders (VAEs), generative adversarial networks (GANs), diffusion models, energy-based models, and flow based methods. Emphasis is placed on model architecture design, training objectives (such as maximum likelihood estimation, adversarial losses, and denoising losses), and inference methods including autoregressive sampling, Langevin dynamics, and Markov chain Monte Carlo (MCMC). Additional topics include multimodal generation, representation learning, foundation model pre-training and post-training, alignments, evaluation techniques, and applications.

Prerequisites:
Students should have prior coursework in machine learning at the level of CS 446 or equivalent, along with a strong background in probability and statistics. Mathematical maturity is expected, including familiarity with advanced linear algebra, multivariable calculus, and abstract probabilistic concepts. Experience with Python programming and numerical libraries such as NumPy and PyTorch is required. Prior exposure to generative models is helpful but not necessary.

Class Meeting Time:
Mon, Wed, 9:30am–10:45pm, SC 0216.
Lectures will be conducted in person and via zoom. The zoom session will be recorded.

Instructor: Prof. Tong Zhang

• Email: tozhang@illinois.edu

• Office Hour: Mon 11:00am – 12:00pm

• Office Location: SC 2118 or via zoom

Teaching Assistant

• Shivanshu Shekhar

• Email: shekhar6@illinois.edu

• Office Hour: Friday, 12:00PM – 1:00PM

• Office Location: SC 1112 or via: zoom.

Course Material:

• Lecture slides (distributed before each lecture)

• Reference papers

Course Resources:

• Website: https://courses.grainger.illinois.edu/cs498tzu/sp2026/

• Canvas: https://canvas.illinois.edu/courses/65164

• Gradescope: can be accessed from canvas

Course Credit:

• Undergraduate: 3

• Graduate: 3 or 4

Grading:

• 3-credit: best four out of six homeworks (55%) + in class exams (45%)

• 4-credit: six homeworks (55%) + in class exams (45%)

Lecture Topics (tentative)

• Lecture 01 (Jan 21): Introduction

Part I: Review of basic neural architectures, optimization, representation learning

• lecture 02 (Jan 26): fully connected neural networks

• lecture 03 (Jan 28): stochastic optimization

• lecture 04 (Feb 2): convolutional neural networks

• home work #1

• lecture 05 (Feb 4): neural sequence models

• lecture 06 (Feb 9): encoder, decoder, and autoencoder

• lecture 07 (Feb 11): variational autoencoder

• home work #2

• lecture 08 (Feb 16): contrastive learning

• lecture 09 (Feb 18): Part I in class exam

Part II: Large Language Model Text Generation

• lecture 09 (Feb 23): transformer

• lecture 11 (Feb 25): encoder only, encoder-decoder, and decoder only transformer models

• lecture 12 (March 2): advanced transformer designs

• home work #3

• lecture 13 (March 4): instruction tuning

• lecture 14 (March 9): scaling law, cot, and emergent abilities

• lecture 15 (March 11): resource efficient training

• home work #4

• spring break

• lecture 16 (March 23): reinforcement learning with human feedback

• lecture 17 (March 25): reasoning and verifiable rewards

• lecture 18 (March 30): evaluation

• lecture 19 (April 1): Part II in class exam

Part III: Multimodal Model and Image Generation

• lecture 20 (April 6): visual language model

• lecture 21 (April 8): auto-regressive image generation

• lecture 22 (April 13): energy based model

• home work #5

• lecture 23 (April 15): MCMC and Langevin sampling

• lecture 24 (April 20): generative adversarial network

• lecture 25 (April 22): flow matching

• home work #6

• lecture 26 (April 27): Unet and diffusion transformer

• lecture 27 (April 29): score matching

• lecture 28 (May 4): diffusion model

• lecture 29 (May 6): Part III in class exam