Northern Illinois University Computer Science Faculty Talks

Event Sponsor: 
Argonne Leadership Computing Facility Seminar
Start Date: 
Nov 8 2019 - 10:00am
Building 241/Room D172
Argonne National Laboratory
Pratool Bharti, Assistant Professor
David Koop, Assistant Professor
Maoyuan Sun, Assistant Professor
Michael Papka

Context-aware Machine Learning Models for Personalized and Public Health


Pratool Bharti, Assistant Professor

Innovations in designing Machine Learning models (and creating novel applications) have been growing at a rapid pace in the last decade. More recently, there is an earnest interest in context-aware learning, where the goal is to carefully extract limited contexts within the domain of interest during model development and execution. Such contexts can be many and also domain specific—for example, physiology of the human body in classifying physical activities; evolution of anatomies in classifying biological species; dynamically changing outdoor environments during modeling and optimization of transportation systems; and much more. The motivation for integrating contexts during learning not only improves accuracy, but also saves on implementation/ execution cost. However, this process is challenging and often requires multi-disciplinary expertise.

In this talk, I will highlight my recent technical contributions in this space. First, I will present my research on designing physiology-aware learning models to accurately classify complex human activities using wearable devices that are significant for personalized elder care. The innovation here is careful integration of multi-modal inertial sensory data from multiple wearable devices emplaced across multiple positions in the human body, and finally integrating human physiology into decision making. Second, I will present my results on neural network models to classify genus and species types of mosquitoes from smart generated images taken by experts or by ordinary citizens. The innovation here lies in extracting contextually relevant anatomies (e.g., head, thorax, wings, legs) from mosquito images, and assigning appropriate weights to only the most critical anatomical component(s) for accurate classification. This work is expected to have significant impact in automating mosquito surveillance and related public health efforts in the US and across the globe. I will also briefly explain my work on vision-guided automation of freight transportation under challenging environmental conditions and dynamics.

Supporting Reproducible Exploratory Data Analysis

David Koop, Assistant Professor

As computational data analysis and visualization tools permeate more fields, the ability to generate results has often outpaced the ability to meaningfully record and reflect on them. Interactive visual applications on the web allow users to explore data and gain insight, but these findings are often ephemeral, lost when a new page is loaded. Computational notebooks combine cells of code with their outputs and explanatory text, and have become key tools in exploratory data analysis. However, the ability to modify and reorder cells often produces unknown or ambiguous dependencies, leading to problems with reproducing past results. In both settings, being able to rapidly explore data seems to be in tension with being able to reproduce and extend that work. My research aims to develop methods that improve the reproducibility of interactive web applications and computational notebooks without impeding the fast and flexible analyses these environments facilitate. Furthermore, making individual explorations more robust and reproducible can often enhance collaboration, allowing others to understand and extend that work.

Engaging Data with Visual Analytics

Maoyuan Sun, Assistant Professor

We are in the midst of a data deluge that shows no signs of slowing down. Human behavior is increasingly mediated by technology that captures, stores, and transmits data at speeds and scales never before imagined. In such an increasingly data-rich world, enabling humans to make sense of big data offers us almost boundless possibilities to learn about ourselves and improve the world. In this talk, I will introduce visual analytics, an interdisciplinary field that marries information visualization with data mining / machine learning and present my work on supporting sensemaking of complex relationships. As a concrete example of involving human in the loop for data analytics, I will go through my work of visual analytics with biclusters. I will discuss the concept and design space of biclusters and present several visualization designs of interactive biclusters. I will also briefly present a few of my other on-going projects and discuss my future research plans.