Project title: Enabling Real-time, Scalable and Secure Collaborative Intelligence on the Edge

Investigators: Zheng Dong, Weisong Shi

Abstract: With the proliferation of embedded systems, multicore computing devices enable the recent trend of moving computation from the centralized cloud to distributed edge platforms. This trend yields new products and services across smart infrastructures in smart cities. However, as real-time workloads are executed at the edge computing platforms, the performance bottleneck is transferred from the edge-cloud communication to on-chip communication. The system’s real-time performance faces new system-architectural challenges for the Network-On-Chip (NoC), which are scalability and security. These challenges are hinged with dynamic data distributions across different users. This project aims to design a real-time and scalable NoC for implementing real-time collaborative learning algorithms. The key strategy is to orchestrate a system-architecture and algorithm co-design to explore the new design space on the edge computing platform.

To cope with the research challenges, a comprehensive architecture will be developed to address these multifaceted problems through a hardware and software co-design, which consists of three key thrusts: (i) designing an interconnect, which will eliminate non-predictability barrier on the NoC; (ii) establishing a scalable virtualized transaction environment for the collaborative learning system to guarantee that all the real-time transaction tasks can complete at the right time; (iii) implementing a real-time and secure multi-target tracking system on the edge platform in light of the newly proposed architecture. The proposed research will be evaluated using the physical platform Equinox, with indoor and outdoor studies beyond simulation.

This research will open a new dimension of research and educational opportunities. In particular, the success of the project will provide a hardware/software package that can enhance the real-time collaborative computing on the edge. The resulted interconnect and Equinox are ready-to-use platforms that will allow experts/researchers to easily examine their research designs regarding collaborative learning and real-time edge computing, thereby sealing the gap between different research fields. Educational efforts will be devoted to (i) curriculum design for the undergraduate and graduate program, (ii) summer camp development for middle and high school students, and teachers, (iii) broadening participation in computing and engineering, at the Wayne State University.

Project title: Neural Conversational Agent for Automated Weight Loss Counseling

Investigators: Alexander Kotov, April Idalski Carcone, and Elizabeth Towner

Abstract: Obesity is one of the most important medical and public health problems in the United States. According to recent nationally representative studies, every third adult in the U.S. is obese. Motivational Interviewing (MI), a client-centered and directive approach to behavior change counseling, has been widely adapted for treating obesity. Despite the evidence presented in published meta-reviews that suggests that MI is effective at activating behavioral changes, anthropometric changes are less significant. At the same time, there are several access barriers to this type of behavioral health care, such as shortage of human counselors in certain geographical areas, long wait times, cost, and fear of judgment. Recent advances in deep learning have allowed artificial intelligence (AI) methods to expand into the areas of health care that were previously thought to be the exclusive province of human experts, such as clinical diagnostics.

Behavioral health and MI, however, are the areas of medicine that have not yet substantially benefitted from modern AI technologies, such as neural conversational agents. To address this limitation, the proposed project aims to test the feasibility and usability of using neural conversational agents for automated behavioral counseling with a focus on weight loss. Specifically, we build on recent advances in deep learning, such as conversational agents, neural attention, transformers, supervised policy learning, variational autoencoders and adversarial training, and aim to develop and validate Neural Agent for Obesity Motivational Interviewing (NAOMI), a mobile device (smartphone or tablet) application to conduct automated MI counseling focused on weight loss. NAOMI is based on a novel neural architecture, which consists of neural networks that can be independently and collectively trained using the proposed multi-stage procedure to learn communication behaviors, which should be strategically utilized during different stages of an MI counseling session depending on the observed interactions and generate responses that are grounded in session context and reflect patient’s language.

We will recruit 40 obese adults, who will interact with NAOMI and provide their feedback through semi-structured qualitative interviews. We plan on conducting at most 4 iterative development cycles of NAOMI with 10 patients participating in each cycle. We will conduct a mixed-methods sub-study after each development cycle. Quantitative evaluation of NAOMI’s MI counseling skills will be conducted based on the transcripts of participants’ interactions by a coder trained in using the MI Treatment Integrity (MITI) coding system, a standard instrument for assessing MI fidelity. Qualitative interviews with the participants will be analyzed using Framework Matrix Analysis. The methods and techniques proposed in this project can be adapted to other types of psychotherapeutic interventions besides MI and to other conditions besides obesity.

Project title: CRII: III: Learning to Integrate Heterogeneous Data from Disparate Sources for Disease Subtyping

Investigators: Suzan Arslanturk

Abstract: Patients suffering from the same cancer disease often not only experience a high degree of symptomatic variability but also respond significantly different to the same treatment. As a result, many cancers are over-diagnosed causing patients to receive unnecessary and costly cancer treatments, while some patients do not receive the needed treatment. These can be greatly reduced with targeted treatments that result in greater efficacy and fewer debilitating or dose limiting side effects. Hence, the discovery of patient subgroups and/or disease subtypes that differ in cancer progression will help better tailor treatments with reduced lethality and improve the quality of life by avoiding over-treatment. Studies have shown that, in addition to the genetic factors, such factors as clinical history, hormonal exposure, lifestyle, and epidemiologic factors may play an important role in the onset and progression of cancers. Hence, investigation of clinically relevant disease subtypes cannot be achieved by solely analyzing data from a single source, i.e. only the genetic composition. This project will build innovative machine learning technologies to integrate knowledge collected from multiple repositories from different cohorts of multiple data types (e.g. genomic, clinical, and epidemiologic) for a more accurate and robust discovery of disease subgroups and/or patient subtypes. Its activities will thus advance the field of multi-type data integration from disparate sources with different cohorts. Anticipated results will greatly benefit the society by reducing the health care costs and improving patient care by distinguishing between patients who are at higher-risk and need the most aggressive treatments from those who will never progress, recur, or develop resistance to treatments. The interdisciplinary nature of the project will further the education of undergraduate and graduate students through cross-disciplinary training by bridging the fields of computer science, biology, and engineering.

To meet these goals, the project will investigate a novel clinically-relevant disease subtyping system by flexibly integrating collective information available through multiple studies with different cohorts and heterogeneous data types using innovative deep learning models and algorithms. Here, the investigator hypothesizes that if the information in partially coupled data from disparate sources is integrated in a meaningful way that is statistically sound and robust, then the analysis on the integrated data would lead to a more unified picture and global view of the system, and thus, to a more accurate and robust discovery of disease subtypes. Existing data fusion approaches suffer from such challenges as uniqueness, interpretability, high-dimensionality of the feature space and linearity assumptions. This project will develop theory, algorithms, and implementation of a deep machine learning technique capable of discovering the salient knowledge of the learning task during the integration of disparate datasets. Therefore, it is expected to overcome such challenges as high dimensionality of genomic datasets. The project would result in a valuable tool with broad implications and utility in cancer research. These findings will not only provide useful and valuable models to identify patient subgroups and/or disease subtypes, but will also result in a valuable precision medicine resource for the wider scientific community on other diseases.