Current Seed Grants 2024

Graphs are a major part of science, as many computational problems in various fields can be modeled or solved using graphs. In many cases, there are two aspects to the solution. One is the high-level algorithmic aspect, e.g., the mathematical modeling of the problem and the algorithm for its solution, and the other aspect is the infrastructure of the computation, e.g., the efficient (parallel) implementation of the methods on common or dedicated hardware. As everyday problems only get bigger and more complex, there is a growing need for scalable graph algorithms and their efficient implementation. Hence, in this research, we wish to improve the scalability of computational graph primitives and parallel implementation of low-level algorithms, and embed them in graph signal processing tools and graph neural networks for various applications. We will consider mostly GPU hardware, but also CPU and the new Maverick chip in collaboration with NextSilicon (https://www.nextsilicon.com/maverick).

Preference queries leverage different preference aggregation methods to aggregate individual preferences in a systematic manner and come up with a single output  that is most representative. These queries are prevalent in high-fidelity applications, including search, ranking and recommendation, hiring and admission, and electoral voting systems. In many such applications the preference queries are perpetual; that is, they need to be aggregated over time. While non-perpetual preference aggregation methods are well studied, this is not the case in the perpetual setting. We plan to explore methods to perpetually aggregate individual preferences in a systematic manner and come up with an output over the time horizon that is most representative. We plan to investigate the computational complexity of these methods and design efficient algorithms to implement them either exactly or approximately. We also plan to design optimal or near optimal algorithms for perpetual preference aggregation that satisfy fairness constraints employing two complementing approaches. In the first approach we are going to define a metric on the possible outcomes of the perpetual preference aggregation and find the outcome that satisfies the fairness constraint and is also closest (with respect to the defined metric) to the best unconstrained outcome. In the second approach we are going to define a metric on the possible inputs and find the smallest amount of change in the input (with respect to the defined metric) that guarantees that the best unconstrained outcome satisfies the fairness constraints.

Recent reports suggest that robots in care and home settings may help battle loneliness in the older population and serve as useful aids in therapy and rehabilitation. While the concept of socially assistive robots holds promise, long-term, sustained use has not yet been achieved. Our goal is to design robots that are adaptive – that understand their human partners better and are affective – to both persistent and transient characteristics of the person (e.g., attachment styles and moods) with the aim of achieving better interactions between people, and assistive agents and sustain engagement levels. Clinical studies and self-reports can offer valuable insights regarding behavioral, functional, psychological, and social changes in users, but they are often limited by cost and scale. Through this project, we aim to explore the feasibility of using multimodal wearable sensors as an alternative that can alleviate these shortcomings through continuous, objective, real-time measurements of user progress while interacting with these robots. We aim to develop a multi-physiological sensing framework for real-time cognitive and microexpression detection, specifically, to infer users’ attachment styles, and conduct user studies to study the impact of spatio-social-awareness on long-term affective interactions. The technical solutions we propose in this work fall broadly under the topics of: (1) Affective Computing and Human-Robot-Interaction, (2) Mobile and Pervasive Computing, and (3) Artificial Intelligence of Things.

MicroRNAs (miRNAs) are small non-coding RNAs involved in post-transcriptional gene regulation, impacting diverse biological processes and disease pathogenesis. Despite their importance, accurately identifying miRNA promoter regions and distinguishing genuine precursor miRNAs from genomic hairpin-like structures remain challenging. Leveraging Large Language Models (LLMs) like GPT and BERT offers promise in modeling miRNA sequences and structures, but their application to genomics faces difficulties due to differences in data structure and tokenization. Additionally, existing explainability tools for LLMs are not suited for genomic data analysis. This proposal seeks to address these challenges by adapting LLM algorithms and explainability tools to genomic contexts, focusing on miRNA promoter classification and pre-miRNA identification. The goal is to enhance prediction accuracy and provide tailored explainability outputs for genomic language analysis.

Our research addresses realistic distributed applications, in which humans and technology interact and aim to optimize mutual goals (e.g., IoT applications). These include, among others, device scheduling in smart homes, target tracking in sensor networks, and disaster response. All these include multiple events that require service from a team of agents.

Recently, the Service-Oriented Multi-Agent Optimization Problems (SOMAOP) model was proposed for representing such scenarios, along with algorithms for solving them. Unlike former approaches, in SOMAOP, there are two types of agents: service-requesting agents and service-supplying agents. SOMAOP-solving algorithms manipulate auction and matching methods to achieve high-quality schedules for service providers. However, the SOMAOP model did not include means for representing uncertainty elements, which are common in such scenarios. In this research, we will extend the SOMAOP model so that such uncertain elements can be represented, and we will propose algorithms for solving the extended model.