2022 Graduate Student Research Poster Room

Our spring 2022 event featured a virtual Poster Room.
We encourage you to visit the various areas of research that our students have been exploring.

Research Areas

Applied Electromagnetics and Acoustics

Communications, Networks, Privacy, and Security

Computer Systems and Architecture

Energy Systems

Machine Learning, Signal Processing, and Info. Theory

Optics and Photonics

Optimization and Control

Solid-state Electronics and Quantum Technologies

Research Poster: Mitigating Heat Stress in Dairy Cattle using a Physiological Sensing Behavior Analysis Microclimate Control Loop
Poster # 1

Mitigating Heat Stress in Dairy Cattle using a Physiological Sensing-Behavior Analysis-Microclimate Control Loop

Hien Vu and Omkar Prabhune (Advisor: Younghyun Kim)

Heat stress is one of the main factors that reduce the productivity of dairy cattle. The dairy industry has been using a large amount of electricity and water just to keep the cattle comfortable in hot weather. Our goal is to design and implement a cyber-physical system (CPS) that reduces heat stress in dairy cattle with high energy and water efficiency. By collecting various environmental parameters and behavioral features of the cattle under heat stress, we can create new psychological models for understanding cattle behaviors and implement a control system to reduce heat stress in dairy cattle.

Research Poster: Ultra Low-Power Machine Learning at the Edge
Poster #2

Ultra Low-Power Machine Learning at the Edge

Tianen Chen and Setareh Behroozi (Advisor: Younghyun Kim)

Smart devices with emerging applications such as recognition tasks for natural user interfaces like voice and gesture demand execution of machine learning (ML) algorithms with specially designed neural networks (NNs) on the edge. Executing part of the task on the edge contributes to less network load, reduced latency, and enhanced privacy. However, the adoption of NNs on the edge is hindered since edge devices contain limited computation resources and are powered by battery or energy harvesting schemes. With an upward trend in NN model size, conventional load and compute architecture–neural professing element (NPE) suffers more than before from long latency and high energy consumption of memory accesses. A promising solution to remove the expensive memory accesses is logic-based NNs, where neurons are implemented as a Boolean logic circuit rather than computed on NPEs. The main challenges in realizing the energy efficiency advantages of logic-based NNs are, first, lack of scalability because of large sizes of logic circuits and, second, inflexibility due to hard-wired weights. The proposed architecture addresses these challenges by leveraging the benefit of logic-based NNs in energy efficiency and reconfigurability of NPE-based NNs. We propose i) logic minimization method that exploits the error resilience of NNs to reduce the logic size, and ii) heterogeneous logic-NPE architecture for reconfigurability.

Research Poster: Improving Fairness via Federated Learning
Poster #3

Improving Fairness via Federated Learning

Yuchen Zeng and Hongxu Chen (Advisor: Kangwook Lee)

Recently, lots of algorithms have been proposed for learning a fair classifier from decentralized data. However, many theoretical and algorithmic questions remain open. First, is federated learning necessary, i.e., can we simply train locally fair classifiers and aggregate them? In this work, we first propose a new theoretical framework, with which we demonstrate that federated learning can strictly boost model fairness compared with such non-federated algorithms. We then theoretically and empirically show that the performance tradeoff of FedAvg-based fair learning algorithms is strictly worse than that of a fair classifier trained on centralized data. To bridge this gap, we propose FedFB, a private fair learning algorithm on decentralized data. The key idea is to modify the FedAvg protocol so that it can effectively mimic the centralized fair learning. Our experimental results show that FedFB significantly outperforms existing approaches, sometimes matching the performance of the centrally trained model.

Research Poster: An Exponential Improvement on the Memorization Capacity of Deep Threshold Networks
Poster #4

An Exponential Improvement on the Memorization Capacity of Deep Threshold Networks

Shashank Rajput and Kartik Sreenivasan (Advisors: Dimitris Papailiopoulos and Amin Karbasi)

It is well known that modern deep neural networks are powerful enough to memorize datasets even when the labels have been randomized. Recently, Vershynin (2020) settled a long standing question by Baum (1988), proving that deep threshold networks can memorize n points in d dimensions using O(e^(1/δ^2)+√n) neurons and O(e^(1/δ^2)(d+√n)+n) weights, where δ is the minimum distance between the points. In this work, we improve the dependence on δ from exponential to almost linear, proving that O(1/δ+√n) neurons and O(d/δ+n) weights are sufficient. Our construction uses Gaussian random weights only in the first layer, while all the subsequent layers use binary or integer weights. We also prove new lower bounds by connecting memorization in neural networks to the purely geometric problem of separating n points on a sphere using hyperplanes.

Research Poster: GenLabel: Mixup Relabeling using Generative Models
Poster #5

GenLabel: Mixup Relabeling using Generative Models

Jy-yong Sohn, Liang Shang, Hongxu Chen and Jaekyun Moon (Advisors: Dimitris Papailiopoulos and Kangwook Lee)

Mixup is a data augmentation method that generates new data points by mixing a pair of input data. While mixup generally improves the prediction performance, it sometimes degrades the performance. In this paper, we first identify the main causes of this phenomenon by theoretically and empirically analyzing the mixup algorithm. To resolve this, we propose GenLabel, a simple yet effective relabeling algorithm designed for mixup. In particular, GenLabel helps the mixup algorithm correctly label mixup samples by learning the class-conditional data distribution using generative models. Via theoretical and empirical analysis, we show that mixup, when used together with GenLabel, can effectively resolve the aforementioned phenomenon, improving the accuracy of mixup-trained model.

Research Poster for "Design of high speed bearingless generator for combined heat and power plant."
Poster #6

Design of High Speed Bearingless Generator for Combined Heat and Power Plant

Imthiaz Ahmed (Advisor: Eric Severson)

1. High-speed generators are gaining popularity due to their compact design, high power density, and high efficiency.
2. Lifetime and reliability of these generators are often restricted by mechanical bearings.
3. Bearingless generators offer an alternate solution where the rotor is levitated using suspension forces created by magnetic fields in the generator’s air gap. This significantly improves the life of high-speed generators.

Research Poster: Breaking Fair Binary Classification with Optimal Flipping Attacks
Poster #7

Breaking Fair Binary Classification with Optimal Flipping Attacks

Changhun Jo and Jy-yong Sohn (Advisor: Kangwook Lee)

Minimizing risk with fairness constraints is one of the popular approaches to learning a fair classifier. Recent works showed that this approach yields an unfair classifier if the training set is corrupted. In this work, we characterize the minimum amount of data corruption required for a successful flipping attack. We find a lower bound on this quantity and propose an upper bound by constructing an explicit data poisoning algorithm. Our bounds are tight when the target model is the unconstrained risk minimizer. Inspired by the attack strategy designed for the upper bound, we propose an efficient sensitive attribute flipping attack algorithm that can compromise the performance of fair learning algorithms.

Poster #8

Acoustic Topological Insulators: Design, Fabrication and Characterization

Michael Wang (Advisor: Prof. Chu Ma)

With the development of metamaterials, acoustic topological insulators (TIs) have aroused great academic interest in the field of biology, noise canceling, computation, etc. As a novel structure, a honeycomb lattice, also named phononic ‘graphene’, has been introduced to guide sound waves traveling along a designated direction. Our simulation has shown that honeycomb lattice made of pure metallic or core-shell metal-plastic pillars forms acoustic double Dirac cone edge state by tuning the radius of pillars due to the large impedance contrast of these materials. The edge state helps realize MHz waveguide under water. Phononic ‘graphene’ Tis can be extended to explore the application of ultrasonic quantum behavior in quantum computing field.

Poster #9

6.78MHz GaN Inverter for Capacitive Power Transfer Field Excitation in Synchronous Machines

Marisa Tisler (Advisor: Prof. Daniel Ludois)

Wound field synchronous machines (WFSMs) are an attractive alternative to permanent magnet synchronous machines (PMSMs) given their competitive power/torque density and independence from volatile prices of rare earth elements in common permanent magnets. However, WFSMs typically suffer from reliability issues with mechanical-contact carbon brushes, or the added complexity of inductive brushless exciters required for DC excitation of the rotating field winding. A full-bridge soft-switching GaN inverter is developed with a series resonant system to transfer power through the electrical capacitance between electrodes on stationary and rotating printed circuit boards attached to the machine’s rotor. The current can then be rectified to provide DC power to the rotating field winding; thus, eliminating the need for sliding, mechanical contacts. Only simple printed circuit boards (PCBs) are used that would naturally need to be present for the rotor’s rotating rectifier. The new excitation prototype is a compact system designed to provide rotor excitation in the kilowatt range while operating the soft-switching GaN inverter within the ISM radio band at 6.78MHz.

Poster #10

Analysis and Simulation of High-Speed Integrated Motor Compressor Concept using FSPM topology

Leyue Zhang (Advisor: Prof. Bulent Sarlioglu)

The goal of this project is to continue to investigate and design our integrated motor compressor concept for high-speed operation. By shaping the salient rotor poles of the flux-switching PM machine as airfoils of axial-flow compressor rotor, we can integrate the compressor with the electric machine to achieve higher system efficiency and compactness. So far, the low-speed version of this concept has been designed and investigated. The high-speed machine has its challenges, such as structural strength, high-frequency AC loss, and iron loss. In the design process, both electromagnetic and thermodynamic performances need to be taken into consideration. Especially the blade shape, solidity, blade angles, and angle of attack need to be carefully designed. In this research, first, requirements for the high-speed integrated motor-compressor is being developed. Then, the needs and challenges in the machine design process are also investigated. Analytical analysis and finite element analysis are performed to design the high-speed integrated motor compressor. The future work includes optimizing the preliminary design to achieve better performance, prototyping, and testing the machine.

Poster #11

Rethinking the Control of Bearingless Motors

Nathan Petersen (Advisor: Prof. Eric Severson)

Applying physics-based control techniques to magnetically levitated motor systems enables more insight into the fundamental operation of these machines.

Poster #12

Tunable infrared optics based on phase-transition materials

Chenghao Wan (Advisor: Prof. Mikhail Kats)

This poster summarizes our group’s work on realizing tunable optical devices in the infrared spectral range using phase-transition materials—e.g., vanadium dioxide (VO2) and samarium nickelate (SmNiO3)—whose optical properties are tunable via external stimuli such as temperature, incident intensity, and strain. By incorporating the phase-transition materials into nanophotonic geometries such as metasurfaces or thin-film assemblies, we have demonstrated a variety of tunable functional devices including optical limiters, nonlinear optical isolators, switchable filters, and thermal emitters. In the poster, we also highlight our capabilities in the experimental characterization of these dynamic optical materials, the design and optimization of optical devices using multiphysics simulations, and the fabrication and testing of these devices.

Research Poster: Distribution Grid Optimization- Mitigating Voltage Unbalance due to High Penetrations of Solar PV
Poster #13

Distribution Grid Optimization- Mitigating Voltage Unbalance due to High Penetrations of Solar PV

Kshitij Girigoudar (Advisor: Line Roald)

Today’s distribution grids are receiving lot of research attention since they have been experiencing a significant transformation due to an increasing amount of distributed energy resources like rooftop solar PV. Our research work is focused on using optimization-based methods to address the power quality issue of voltage unbalance due to large-scale penetration of solar PV installations in the distribution grids. By lowering voltage unbalance levels, we can decrease losses in the network and avoid premature failure of three-phase equipment such as induction motor loads connected to the grid.

Research Paper: A Posteriori Chance Constraint Tuning for DC Optimal Power Flow
Poster #14

A Posteriori Chance Constraint Tuning for DC Optimal Power Flow

Ashley Hou (Advisor: Line Roald)

Abstract: We consider a chance-constrained formulation of the optimal power flow problem to handle uncertainties resulting from renewable generation and load variability. We propose a tuning method that iterates between solving an approximated reformulation of the optimization problem and using a posteriori sample-based evaluations to refine the reformulation. Our method does not rely on any distributional assumptions on the uncertainty and is applicable to both single and joint chance constraints and does not rely on any distributional assumptions on the uncertainty. In a case study for the IEEE 24-bus system, we demonstrate that our method is computationally efficient and enforces chance constraints without over-conservatism.

Research Poster: Efficient Real-time Video Stabilization with a Novel Least Squares Formulation
Poster #15

Efficient Real-time Video Stabilization with a Novel Least Squares Formulation

Jianwei Ke, Alex Watras, Jae-Jun Kim and Hewei Liu (Advisors: Hongrui Jiang and Yu Hen Hu)

We present a novel video stabilization algorithm (LSstab) that removes unwanted motions in real-time. LSstab is based on a novel least squares formulation of the smoothing cost function to alleviate the undesirable camera jitter. A recursive least square solver is derived to minimize the smoothing cost function with an O(N) computation complexity. LSstab is evaluated using a suite of publicly available videos against the state of the art video stabilization methods. Results show LSstab reaches comparable or better performance, achieving real-time processing speed when a GPU is used.

Research Paper: Linear Motor Design to Electrify the Hydraulic Charge Pump of Off-Highway Vehicles
Poster #16

Linear Motor Design to Electrify the Hydraulic Charge Pump of Off-Highway Vehicles

Anvar Khamitov (Advisor: Eric L Severson)

Off-highway vehicles (tractors, excavators, wheel loaders, etc.) use an internal combustion engine to drive a hydraulic pump that distributes power through hydraulic fluid to remote actuators. These systems are notoriously inefficient (only 20% of engine shaft power makes it to the implements). This project targets a step increase in the efficiency of off-highway vehicles through partial electrification. Specifically, the project is focused on developing an integrated electric motor-hydraulic pump to replace the inefficient charge pump central to these systems. Existing attempts at electrifying hydraulic systems use distinct modular components coupled together. This results in redundant bearings, interfaces, shafts, and seals and creates a system with only a modest energy efficiency and large rotational inertia resulting in poor dynamic response. This project takes a fundamentally different approach by developing a single machine that re-uses surfaces of a linear electric motor to function as a piston pump. The two ends of the motor mover serve as pistons, check valves control the fluid flow from a low-pressure tank to a high-pressure outlet, and mechanical springs allow the machine to operate efficiently at a resonant frequency. The linear motor design was optimized using a multi-objective optimization algorithm linked to an FEA model of the motor for a total system efficiency of 80%. A prototype based around a single-phase double-sided permanent magnet linear motor was fabricated and validated experimentally. The closed loop position control was designed, integrated motor-pump system has been built (with collaborators at the University of Minnesota) and is currently under experimental testing.