Events Listing

Cloud-based Field-Programmable Gate Arrays (FPGAs) are widely used by a broad range of users. These devices often perform highly sensitive computations that result in data that must be secured. In this talk, the state-of-the-art in cloud FPGA security is reviewed with a focus on voltage-based attacks and remediations. Our research shows that data stored in FPGA-attached DRAM remains viable for nearly 20 minutes even when data refresh is not performed and the data is not directly accessed. This issue is a significant concern when multiple users consecutively share an FPGA platform and DRAM data is not explicitly cleared after an initial user stops using the FPGA. We demonstrate the threat and describe remediations using results collected from the Open Cloud Testbed, an open cloud architecture which includes FPGAs. A use of FPGA-equipped edge cloud involving data offloading from drones is also outlined to show the need for secure cloud computing. 

The impact of artificial intelligence (AI) is rapidly growing and is increasingly pivotal across a wide range of disciplines, from innovative scientific research to practical, everyday applications. The powerful capabilities of AI—spanning data analysis, predictive modeling, and beyond—equip researchers and professionals with unparalleled tools to tackle complex problems, push the boundaries of scientific discovery, and elevate productivity to unprecedented levels. This talk will explore the integration of AI in diagnosing motor faults and advancing motor design, highlighting how AI can significantly enhance the reliability and performance of electric motors in green transportation. It will delve into the use of machine learning and deep learning models to predict and prevent motor failures (e.g., inter-turn short-circuits, demagnetization, and bearing faults), which is essential for ensuring safety and reliability in transportation and industry. Furthermore, the talk will highlight AI-driven innovations in motor design, such as noise-reduction, offering insights into how AI can revolutionize traditional motor systems and contribute to ongoing improvements in predictive maintenance and design practices.

Fragility phenomena appear to be more ubiquitous than is generally believed. Hydrodynamic stability of wall-bounded shear flows is one striking example where flow fluctuations are highly susceptible to even small irregularities. The well-known phenomenon of Anderson localization is another one where small amounts of material disorder can radically alter the nature of dynamical modes. Recent work has shown that such localization phenomena can also occur without the presence of medium disorder, but rather due to complex geometry such as in vibrational modes of several naturally occurring biomolecules. The common thread between these seemingly disparate phenomena is the fragility of eigenvalues/vectors of large-scale matrices and operators when perturbed in various ways. 

Reliable AI by Design

AI models are increasingly being deployed in domains where violating a safety, physical, or structural constraint is not merely undesirable, but unacceptable. Yet most modern learning systems encode such requirements through data or soft penalties, providing little assurance when deployed beyond the training distribution. In this talk, I will discuss recent efforts by my group toward building hard constraints directly into learning-based decision-making systems, while retaining the flexibility and scalability that make modern machine learning effective. I will begin with hard-constrained neural networks (HardNets), which embed differentiable constraint-enforcement mechanisms into existing architectures, enabling end-to-end learning under various constraints while preserving expressivity and guaranteeing feasibility. I will then discuss how HardNets can be used to address important problems in dynamical systems and control, including learning chaotic dynamics with boundedness guarantees via energy-constrained operators, as well as safe yet scalable learning-based control and reinforcement learning. Finally, I will turn to modern generative models, namely diffusion and flow-matching models, where treating sampling as an optimal control problem allows us to impose constraints on the final generated sample without unnecessarily restricting the intermediate sampling dynamics, with applications spanning robotic manipulation and planning, PDE control, and language-guided image editing. Together, these advances pave the way for reliable decision-making systems with intrinsic guarantees.

Related papers:

• HardNet: Hard-Constrained Neural Networks with Universal Approximation Guarantees 
• HardNet++: Nonlinear Constraint Enforcement in Neural Networks
• ECO: Energy-Constrained Operator Learning for Chaotic Dynamics with Boundedness Guarantees
• PS2-RL: Provably Safe, Yet Scalable Reinforcement Learning 
• HardFlow: Hard-Constrained Sampling for Flow-Matching Models via Trajectory Optimization
• DiRecT: Safe Diffusion-Based Planning via Receding-Horizon Denoising 

RISC-V Vectors Present and Future: Can they measure up?

The RISC-V CPU core open architecture specification has generated a multitude of microarchitecture implementations predominantly deeply embedded in Systems on Chip (SoC). Recently RISC-V products and research implementation have emerged that include vector units for acceleration. Currently RISC-V vector CPU systems are available commercially, and there is an active body of research designing high performance vector microarchitectures integrated with scalar processing cores.

In this talk, we evaluate performance of existing and proposed RISC-V vector systems using a collection of measurement techniques. We'll present findings from performance measurement of several commercial RISC-V Vector systems. To evaluate proposed systems, we employ a collection of simulation tools from compiler-oriented to FPGA to AI-assisted. 

Artificial Intelligence for research productivity

Monthly Workshop for Graduate Students.
Faculty and undergraduate students are welcome!

This month's talk is by Professor Jarvis Haupt

Large language models were once synonymous with massive cloud infrastructure and specialized data centers. Today, advances in model architecture, inference engines, and quantization techniques have made it possible to run capable AI systems entirely on consumer hardware, including laptops, desktops, and even mobile devices. This talk explores the technological developments that enabled the rise of local AI models and examines the practical tradeoffs involved with running them. After a brief overview of modern transformer-based language models and the computational resources required for inference, we will discuss the central challenge of local deployment: fitting increasingly capable models within limited memory and compute budgets. Particular attention will be given to quantization methods and the GGUF model format, which have become foundational technologies for efficient local inference. We will survey the current ecosystem of tools for self-hosting and interacting with local models, including popular inference frameworks and desktop applications. Along the way, we will discuss hardware considerations, performance expectations, privacy implications, and opportunities for integrating local AI into research and educational activities.

Organic semiconductors are conjugated molecular materials with highly tunable electrical and optical functionality. These materials have found wide interest as thin film components in optoelectronic and photovoltaic devices, where attractive physical properties are combined with high throughput processing on mechanically flexible substrates enabling novel device form factors. To date, the most successful application of organic semiconductor thin films has been in displays based on organic light-emitting devices (OLEDs). An OLED consists of a vertical thin film stack deposited via a high vacuum, physical vapor deposition process. While these layers are typically amorphous and glassy, ongoing work has revealed the complexity and tunability in property and performance that can come with active engineering of molecular orientation.

Central to this talk is how changes in thin film molecular orientation lead to corresponding changes in the associated transition dipole moment (TDM) and permanent dipole moment (PDM) orientations. The TDM orientation plays a role in determining thin film phenomena including birefringence and OLED efficiency, and manipulation of this parameter is already an active part of display design. This talk will focus more heavily on the less widely investigated case of preferential PDM orientation (termed spontaneous orientation polarization, SOP), its impact on charge and exciton behavior in OLEDs, and ultimately its impact on efficiency and operational lifetime. Emphasis will also be placed on discussing means to manipulate PDM orientation via choice of thin film processing conditions and molecular blending, as well as showing how device architecture can be used to mitigate the impact of SOP on OLED performance. The talk will conclude with a discussion of how tuning of molecular orientation in organic thin films represents a largely open axis for engineering the behavior of this important
materials class.

Celebrating our graduating undergraduate students before their CSE commencement ceremony!