AURA 2022 Projects

Marie Merci Allelua


Project title: Understanding and Modeling Gig Workers' Interactions with Instacart

Faculty Mentor: Nikola Banovic

Description: AI has started to transform the nature of work in many sectors of the economy. One of the most tangible transformations has been in the on-demand economy, for services such as grocery delivery, ride-hailing, and other last-mile services, where its advances have allowed a shift towards greater efficiency, through the use of AI-mediated platforms. On-demand work, with its promises of flexibility, independence and entrepreneurship is also an attractive option for individuals seeking a low-barrier entry into employment and economic opportunities. However, several recent debates around the employment status of workers with services such as Uber, Lyft and Instacart have shined a light on the adversarial relationships between gig workers and platforms, and the negative effects of opaque algorithms on workers’ well-being. In this project, we seek to design computational methods to audit these opaque platforms to uncover sources of adversarial human-AI interactions that may be potentially harmful to on-demand workers. Our goal is to understand the design of algorithmic platforms that enhance worker well-being and their access to economic opportunities.

Gloria Berimana


Project title: Computing on encrypted data

Faculty Mentor: Todd Austin

Description: In the age of big data, privacy is a key concern in sharing data. Unfortunately, the field of security is riddled with stories of security attacks… ¦even to the most secure enclaves. The solution we want to investigate with this project uses encryption technology to encrypt data locally, transfer it to the cloud for any required computation, and receive encrypted results back. The enhanced cloud system performs the computation directly on the encrypted data without an access key -- it never accesses the plaintext data nor can it decrypt the sensitive data. Only the end device can decrypt the result and store it locally.

Milkiyas Gebru Gebremichael & Kidus Yoseph Wondimagegnehu


Project title: Accelerating Explainable AI

Faculty Mentor: Valeria Bertacco

Description: Machine learning algorithms are being adopted in a growing number of applications. Some of these applications have critical impacts in social and human aspects, thus there is a pressing need for AI predictions to be explainable, that is, justifiable to end-users. Some algorithms are more amenable to explainability, thus adhering to new legislation, and leading to new task-specific insights. However, their execution through software applications running on multicore processors does not provide results at the pace needed by their users: for some applications (e.g., medical diagnoses), models may take hours, days, or weeks to train, while for others (e.g., autonomous vehicles) decisions are required within milli-seconds. Such demands lead to the need for specialized hardware solutions in this space. This project focuses on the development of specialized architectures and/or memory hierarchies to address this need.

Ayda Sultan Hassen & Simon Pierre Rusekeza


Project title: Computer Vision for Physical and Functional Understanding

Faculty Mentor: David Fouhey

Description: Our lab is broadly focused on building 3D representations of the world and understanding human/object interaction. Potential projects include learning about: navigating environments, object articulations, commonsense physical properties of objects, and hand grasps. Please look at: http://web.eecs.umich.edu/~fouhey/ for a sense of what projects we’ve done in the past. We will find a specific project based on mutual interest and particular abilities (e.g., stronger systems programming abilities, experience with graphics, etc.).

Meti Adane Bayissa


Project title: Computing Systems for Precision Health

Faculty Mentor: Satish Narayanasamy

Description: Sequencing technology has been far outpacing Moore's law over the last couple of decades, both in terms of throughput and cost. Sequencing can help address a wide range of health issues ranging from detecting pathogens to early cancer diagnosis. To realize their full potential, at UM, we are developing a wide range of computing systems and hardware solutions that address both performance and privacy issues.

Dawit Badeg Melka


Project title: Real Time Liquid Biopsy Exploration for Precision Health

Faculty Mentor: Reetuparna Das

Description: Explore different real time sequencing techniques for early cancer diagnosis.

Bontu Fufa Balcha & Claude Kwizera


Project title: Cross-Cultural Natural Language Processing Models for Health Behavior Communication

Faculty Mentor: Rada Mihalcea

Description: Health communication is a cornerstone of preventative healthcare, addressing behaviors such as weight management, exercise, vaccine intake, smoking cessation and more. Much of the work to date on communication strategies for healthy behaviors have assumed a 'one-size-fits-all' approach, without careful regard to the population group where this strategy is being used. Natural Language Processing (NLP) techniques have the potential to change that by (1) using language analysis on large datasets to understand what "drives" a certain population (e.g., values, interests); and (2) generating communication strategies that are aligned with the population's drives. In this project, we will explore natural language processing techniques for effective communication strategies that explicitly account for the "drives" of the groups of people involved. We will focus on communication for one healthy behavior (e.g., vaccination), and work in close collaboration with public health experts.

Tewodros Mesfin Berhanu & Albert Tuyishime


Project title: Observations and Opportunities in Multi-GPU systems

Faculty Mentor: Ronald Dreslinski

Description: GPUs are one of the predominant accelerators for emerging applications such as recommendation systems, machine learning, graph processing, etc. The performance of successive generations of GPU has so far increased by scaling transistor sizes. However, with the slowing down of Moore’s law, the performance of single-GPU systems has faced an obstacle. As an alternative, recent academic and industry works have proposed multi-GPU systems, where multiple GPUs operate in tandem. Nonetheless, prior works showed that the performance offered from multi-GPU systems is not commensurate with the total computational power of the constituent GPUs. To remedy this, a wide range of architectural optimizations have been proposed. Nonetheless, significant effort is still required to fulfill the remaining performance gap of multi-GPU systems. One of the critical challenges in delivering research works is the absence of a comprehensive benchmark suite for multi-GPU systems. Prior benchmarks are either for single-GPU systems or focus on specific applications/subsystems in multi-GPU systems. On the other hand, the main objective of this work is to deliver a comprehensive benchmark suite for multi-GPU systems. In addition, workload characterization and critical opportunities are pinpointed to facilitate future research works. In order to support the delivery of high quality multi-GPU research, we plan to contribute an open source benchmark suite in the context of multi-GPU systems. In summary, this project plans to deliver the following items: 1) Comprehensive benchmark for multi-GPU systems, 2) Detailed characterization of workloads in multi-GPU systems, 3) Key observations and opportunities in multi-GPU systems.

Natneam Mesele Assefa & Blandine Umuhoza


Project title: Security Project: Stochastic Side Channel Attacks

Faculty Mentor: Todd Austin

Description: Side channel attacks are one of the primary means to attack modern hardware. The attacks use visible properties of a system, like timing, power, and electromagnetics, to infer information about the secrets held within. Eliminating these vulnerabilities is a primary focus of hardware security research today, and the leading approach to stop these attacks is through randomization techniques. For example, to stop cache timing side channels, CPU developers are starting to deploy cache mapping randomization techniques. The goal of this project is to demonstrate that randomization techniques are a futile direction to pursue in stopping side channel attacks. We will do this by developing stochastic side channel attacks, which utilize randomized probes to infer secrets, thus, they are immune to all randomization defenses! Along with these advanced attacks, we will start to work on more durable defenses for hardware side channels.

Sitota Ezra Mersha


Project title: Data-Oblivious Algorithms to Tame the Data Flood

Faculty Mentor: Valeria Bertacco

Description: Data generation is accelerating at a very fast pace worldwide. A large fraction of generated data is processed in data centers, being analyzed to extract relevant information from it. A recent trend in data centers is the creation of “disaggregated memory pools”, that is, data-center nodes comprising primarily memory, which can store a large amount of data efficiently. A key challenge in leveraging these pools efficiently is the high latency of moving data to and from compute nodes to process it. This project explores data-oblivious algorithms to eliminate such high latencies. Data-oblivious algorithms have the key characteristic that the sequence of data accesses they issue is oblivious to the input data on which they are computing. We seek to explore this type of algorithms to optimize the predictability of data transfers, which would in turn reduce or eliminate transfer latencies.

Eyouel Haile Kibret


Project title: Improving end-to-end deep learning recommendation system training by efficient memory and storage utilization

Faculty Mentor: Ronald Dreslinski

Description: Deep learning-based recommendation models are broadly deployed in big technology companies to personalize the experience of their audience. For example, Google uses such models for personalized advertisements, Amazon for recommending items in its catalog, Microsoft for ranking and recommending news to users, Alibaba for recommending products, and Meta for ranking and click-through prediction for news feed and ads.

DLRM requires massive computation and memory resources for training. In DLRM we have dense features that are compute intensive and sparse features that are memory bandwidth and capacity intensive. The sparse features of DLRM require TBs of memory capacity for training. This usually means using multiple training nodes. However, these sparse features have temporal locality that follows power-law distribution. In power law distribution ~80% of data access comes from ~20% of the data. Hence, we can benefit from utilizing the memory hierarchy by optimizing data placement and using sparse feature-aware caching and data movement between different memory hierarchies. Sparse features have irregular memory access patterns. This means with current training systems, even though we benefit from temporal locality, we still waste memory bandwidth because of the irregular memory access. Recent works have shown that we can use storage technologies for DLRM inference. In this project, we first explore caching mechanisms for DLRM using the memory hierarchy with storage technologies. Second, we will explore how we can optimize for irregular memory accesses of DLRM. When using storage technologies, the bandwidth wasted with irregular memory access will even be more problematic. Therefore we want to optimize by implementing efficient sparse feature data reordering mechanisms.

Fekadu Duguma Woltejji


Project title: Design and Manufacturing of Large-Scale Deployable Structures

Faculty Mentor: Evgueni T. Filipov

Description: Deployable structures that use the principles of origami could lead to applications in multiple scales and disciplines from biomedicine to space exploration. In architecture and civil engineering, reconfigurable facades could adapt to the environment, and rapidly deployable shelters and bridges could be used for disaster relief efforts. The objective of this project will be to explore how to scale up principles of origami for structural engineering applications. The student will first create an analytical model to study the motion and geometry of an origami-inspired deployable structure. Next, a laser cutter will be used to fabricate panels for a scaled prototype of the structure. These individual panels will then be interconnected with metallic or plastic hinges that allow for deployment and reconfiguration. The systems will be constructed to minimize the stowed volume while allowing for a reliable deployment that requires minimum force input. Time permitting, the student will conduct experimental testing to quantify the stiffness of different deployable systems.

Jean Paul Habiyakare


Project title: The link between the drinking water distribution systems hydraulics and water quality

Faculty Mentor: Nancy Love

Description: My work is at the interface of drinking water infrastructure and public health. In past work, my collaborators and I (including faculty at AAU) have correlated changes in the microbial composition of drinking water with hydraulic water age in drinking water distribution systems. This work continues in declining population cities in the United States, and is also of interest in rapidly expanding cities in Africa. In both cases, water storage - be it due to stagnant water in a high water age, hydraulically oversized distribution system in a shrinking U.S. city, or due to intermittent water that forces water storage in rapidly expanding cities in Africa - is well known to cause deteriorating water quality. In particular, we are interested in the proliferation of pneumonia-causing bacteria (such as those that cause Legionnaires' Disease) under these scenarios. We hypothesize that pneumonia-causing bacteria outcompete many enteric bacteria under high water age conditions. Our work is evaluating the microbial ecology and chemical conditions in these systems, their propensity to select for more harmful bacteria, the mechanisms of the beneficial growth of undesirable bacteria, and hydraulic conditions that lead to these challenges. The student who works on this can work on either the microbial aspects of the project (lab-based project) or assist with the development of hydraulic models (computation-based project), depending upon their background.

Kalab Yibeltal Assefa


Project title: Automated analog generation

Faculty Mentor: Mehdi Saligane

Description: OpenFASOC stands for open source fully autonomous SoC and has been built on top of OpenROAD for push-button layout generation as part of the current open source effort. This project requires to improve a few of our open-source analog generators (Temperature Sensor, Switched-Cap DC-DC,.. ), which were integrated as part our designs in Google's free shuttles MPW-I / II and our GF12LP tapeout of the OpenTitan SoC which heavily used open source tooling.

Obed Irakoze


Project title: OpenROAD: Open source EDA tooling

Faculty Mentor: Mehdi Saligane

Description: The OpenROAD project's goal is to democratize access to silicon. The students will be assigned to work on helping develop the tools and support other users. Example of tasks would be: 1) work with the developers as AEs to improve the tools, 2) work on enabling new features such as power gating / upf flow, clock gating, etc.., and 3) add additional regression tests.

Emmanuel Mayani


Project title: Integrating more renewable energy into African power grids

Faculty Mentor: Johanna Mathieu

Description: This project will explore the potential advantages and challenges to integrating more distributed renewable energy sources into African power grids. We will develop grid models to investigate how renewable energy impacts grid operations and reliability. We may also explore how technologies like storage and demand response can be used to mitigate challenges that occur when we increase renewable energy penetrations. We will tailor the project to the student's location and interests.

Amir Kelifa Zeinu


Project title: Life Cycle and techno-economic assessments of CO2 utilization technologies

Faculty Mentor: Volker Sick

Description: The project will characterize the distribution of results and recommendations of life cycle and techno-economic assessments for products made from carbon dioxide. The key challenge will be to distinguish between real technical and economic differences and differences that only exist because of the choice of the assessment method. The project will provide a good introduction into what can be made from carbon dioxide, when it makes sense, and why.

Bilen Measho Kidane


Project title: Let’s build some robots

Faculty Mentor: Necmiye Ozay

Description: Open dynamic robot initiative aims to build low cost dynamic robots mostly from 3D printed and off-the-shelf components. The goal of this AURA project is to replicate one or two of the robots proposed as part of this initiative. You will work with a group of 3-4 students and we will provide all the required hardware and components. We have most of the electrical and mechanical systems of the robot ready but it still requires assembling all the pieces and testing it. Once the robots are built, we will work on programming and controlling them to do fun stuff.

Bereket Shimels Ayele


Project title: Formal Verification of a Small Wheeled Robot implementation

Faculty Mentor: Jean-Baptiste Jeannin

Description: Cyber-Physical Systems, physical systems controlled by software, are ubiquitous in daily life, controlling everything from aircraft and automobiles, to appliances and industrial automation. Because many of these systems are employed in safety-critical situations, their operation must be thoroughly certified to ensure that they are error-free. One such method for certifying these systems is formal verification, which involves mathematically proving that programs meet certain conditions for reliability. In our lab, we are working on extending Lustre, an industry-proven programming language for modeling Cyber-Physical Systems, with the ability to formally-verify programs using an enhanced type checker that can prove program properties using refinement types. Named MARVLus (Method for Automated Refinement-type Verification of LUSTRE), this language strives to model cyber-physical systems that can simultaneously be formally verified at compile-time and executable by real robots at run-time. Our current work involves developing the MARVLus compiler and runtime to verify and execute programs intended for a small wheeled robot, and gathering performance metrics. The goal of the AURA project will be to use MARVLus to implement formally verified controllers for a small wheeled robot. Implementing these controllers will lead to interactions with the design team of the language (a PhD student and a few undergraduates), and likely participation in improvements of the language.