Seminars

The proliferation of drones has led to increased security concerns, prompting governments worldwide to develop anti-drone technologies. Recent conflicts, including the Russia-Ukraine and Israel-Hamas wars, have showcased both the deployment of these technologies and the rapid evolution of drones to counter them. This talk presents our research group's comprehensive work on anti-drone technologies, focusing on attacks targeting three critical drone systems: sensing circuits, GPS, and communication channels. We will discuss our "Rocking Drone" [Usenix Security'15] and "…

Optical coherence tomography (OCT) is a technology invented in 1991 to image small critical tissue structures throughout the body with micrometer resolution. It is widely used in the management of eye and coronary heart diseases. In 2023, OCT received broad attention when its inventors received the prestigious Lasker-DeBakey Clinic Medical Research Award and the National Medal of Technology and Innovation from President Biden. For me, it was the culmination of 3 decades of work as an engineer, clinician, and translational researcher, as well as an even longer…

This presentation will explore the critical process of curating medical imaging data for AI algorithm development in ophthalmology, highlighting the challenges and current limitations in data curation. It will discuss benchmark datasets, reference standards for FDA validation, and innovative strategies to enhance data availability. Attendees will gain insights into best practices and future directions in image curation for advancing AI applications in eye care.

Safety, robustness, and interpretability are some of the major challenges in developing systems for human-robot interactions. Learning-from-demonstrations (LfD) is a popular paradigm to obtain effective robot control policies for complex tasks via reinforcement learning without the need to explicitly design reward functions. However, this paradigm typically requires large datasets of user demonstrations. It is also susceptible to imperfections in demonstrations and raises concerns of safety and interpretability in the learned control policies. To address…

In this talk, I will present Imprecise Probabilities (IPs), from their historic philosophical motivations to their applications to frequentist and Bayesian statistics. In turn, I will explore how the statistical approaches to IPs allowed to start the field of Imprecise Probabilistic Machine Learning, and why such a field is of paramount importance in modern ML. I will provide plenty of references, together with some new results on the comparison between Conformal Prediction and Imprecise Probabilistic Machine Learning, and how the latter can be used to derive regions that are narrower than…