<aside>
4 week summer course: July 12 to Aug 7, 2026
Instructor: Prof. Romit Roy Choudhury
Lectures at 4036 CIF, UIUC campus
TAs: 🧓 Debottam Dutta (PhD student) 🧑 Ishaan Gupta (UG)
Lab session or TA office hour (10am to 11:30am)
2 hour lectures per day (1 to 3pm)
Final class grade based on course project
Slack: https://join.slack.com/t/ai-sumer-26/shared_invite/zt-437taou7m-gcDEVPfFfhZYj2Hz2jn1BQ (add your profile picture)
</aside>
<aside>
Download this mini-project and unzip. Please follow the instructions from there.
</aside>
| Lecture | Topic | Material | Video |
|---|---|---|---|
| 1 | Introduction to Generative Models | • Explainer blog PDF | lec1 |
| 2 | Overview of VAE, GAN, Diffusion, Flows. | ||
| Math review: probability (Bayes, MLE) | • Review cheat sheet PDF | lec2 | |
| 3 | Math review: probability (Bayes, MLE, …) | ||
| Math review: linear algebra (Rank, PCA, SVD) | • Review by Kolter PDF | ||
| • PCA notes PDF | lec3 | ||
| 4 | AutoEncoder (AE) and Variational ideas | • VAE blog web | lec4 |
| 5 | Variational AutoEncoders (VAE) | lec5 | |
| 6 | Diffusion models → gentle introduction | lec6 | |
| 7 | Hierarchical ELBO derivation | lec7 | |
| 8 | 3 predictions: image = noise = score function | lec8 | |
| 9 | Guided diffusion (Classifier based guidance) | lec9 | |
| 10 | Classifier free guidance (CFG) | lec10 | |
| 11 | CLIP and T2I models | lec11 | |
| 12 | 🖥️ Guest lecture by Longfei Shangguang | Thursday (July 30) | lec12 |
| 13 | Inverse problems | lec13 | |
| 15 | 🖥️ Guest lecture by Fahim Kawsar | Monday (Aug 3) | lec14 |
| 14 | Inpainting + Low rank Jacobians, NeRFs | lec15 | |
| 16 | Applications and wrap up | lec16 | |
| 17 | Discussion and buffer class |
Dr. Longfei Shangguan (Professor at Univ. of Pittsburgh and Visiting Scientist, Google)
<aside>
Title: From E-Waste to AI Wearables: Low-Cost Health Sensing for Everyone
Abstract: Wearable devices such as Apple Watch and Fitbit wristband allow users to track their health statistics around the clock. They have become increasingly popular over the past few years. However, in the context of underdeveloped regions where people may earn less than 5 US dollars a day, these wearable devices are still pricey and thus constitute a critical bottleneck in their adoption. In this talk, I will present our past and ongoing work on combining signal processing and AI to repurpose low-cost electronic devices, particularly everyday earphones, into health trackers. By learning health-related information from noisy and weak physiological signals, these systems enable applications ranging from heart-rate monitoring and heart-sound recovery to pulse-wave-velocity estimation. I will also share observations and lessons from our pilot study in Senegal, including both the technical and practical challenges of deploying wearable health technologies in resource-constrained communities. Together, this research explores a holistic approach to recycling and repurposing electronic waste while using AI to foster a more sustainable and equitable future for digital health.
Bio: Longfei Shangguan is an Associate Professor at the University of Pittsburgh and a Visiting Faculty Researcher at Google Health. His research interest lies in mobile health, wireless networks, and AIoT. His work has been awarded the Best Paper Nominee at SenSys 2025, Best Paper Award at MobiCom 2024, Best Paper Runner-up Award at MobiCom 2021, Distinguished Paper Award at Ubicomp 2017, and Best Paper Award at TrustCom 2014. He is the recipient of an NSF CAREER Award in 2024, a Google Research Scholar Award in 2023, and the AIoTSys Young Scientist Award in 2023. Longfei earned his Ph.D. from HKUST in 2015 and his bachelor's degree from Xidian University in 2011.
</aside>
Dr. Fahim Kawsar (Professor at Univ. of Glasgow, and CTO at OmniBuds)
<aside>
Biosignals are not physiology—they are incomplete observations of the biological processes that govern human health. As healthcare shifts from episodic diagnosis to continuous prevention, the central challenge is no longer collecting more data, but inferring the latent physiological state that gives rise to these observations.
In this talk, I will present our vision for Physiological Intelligence—a scientific framework that combines next-generation earable sensing with foundation models to continuously infer human physiological state. I begin by demonstrating why the ear is emerging as the next frontier of digital health and how a decade of research—from eSense to OmniBuds and our latest multimodal earable platform—is enabling continuous monitoring of cardiovascular and respiratory physiology, with applications in chronic diseases such as hypertension and COPD.
I will then argue that this is fundamentally a representation problem, not a scaling problem. If biosignals provide only partial observations of the underlying physiological system, simply scaling data or AI models cannot recover the missing physiological information. Instead, physiological AI must first learn to represent temporal dynamics, then preserve biological structure, and finally incorporate clinical knowledge as constraints on inference. Drawing on our recent work, I will show how each step addresses a fundamental limitation of existing physiological AI, moving foundation models from learning biosignals towards representing physiology.
Finally, I will argue that physiology is fundamentally an inverse problem. Because biosignals alone cannot uniquely determine an individual’s physiological state, inference requires additional physiological constraints. In this framework, calibration becomes a source of scientific information rather than an engineering refinement, enabling foundation models to infer latent physiological state rather than directly predict clinical variables.
Ultimately, I will outline a roadmap towards Physiological Intelligence—a new generation of AI built on the premise that biosignals are not the object of learning, but the evidence from which human physiology must be inferred, bringing us closer to continuous, personalized, and truly predictive healthcare.
</aside>