WPI Worcester Polytechnic Institute

Computer Science Department
------------------------------------------

AIRG Topics - Fall 2014

Version: Thu Dec 4 17:07:43 EST 2014

Our group meets on Thursdays at 11:00 a.m.

  • Fall Semester
    • A term FL 320
    • B term Beckett Conf. Rm.
  • Spring Semester
    • C term SL406
    • D term FL311


Dates and topics for this semester are as follows:

Sept 4
A term: FL 320
CS AI Faculty
"AI Research @WPI Overview"
  • Dimitry Berenson
  • Neil Heffernan
  • Mike Gennert
  • Carolina Ruiz

Sept 11
CS AI Faculty
"AI Research @WPI Overview"
  • Dave Brown
  • Dimitry Korkin
  • Sonia Chernova
  • Xiangnan Kong
  • Candace Sidner

Sept 18
Carolina Ruiz
Introduction to Machine Learning and Data Mining
    In this talk, a brief introduction to machine learning and to data mining will be presented. Some terminology (e.g., supervised vs. unsupervised learning) will be introduced, and several machine learning techniques (e.g., artificial neural nets, support vector machines, Bayesian models, clustering) will be described. No previous knowledge of machine learning or data mining will be assumed from the audience.

Sept 25
Ben Suay
Experiences from the DARPA challenge
    In this work, we present our system design, operational procedure, testing process, field results, and lessons learned for the valve-turning task of the DARPA Robotics Challenge (DRC). We present a software framework for cooperative traded control that enables a team of operators to control a remote humanoid robot over an unreliable communication link. Our system, composed of software modules running on-board the robot and on a remote workstation, allows the operators to specify the manipulation task in a straightforward manner. In addition, we have defined a clear operational procedure for the operators to manage the teleoperation task, designed to improve situation awareness and expedite task completion. Our testing process, consisting of hands-on intensive testing, remote testing, and remote practice runs, demonstrates that our framework is able to perform reliably and is resilient to unreliable network conditions. We analyze our approach, field tests, and experience at the DRC Trials and discuss lessons learned which may be useful for others when designing similar systems.

Oct 2
Xiaolu Xiong
Limits to Accuracy: How Well Can We Do at Student Modeling?
    There has been a large body of work in the field of Educational Data Mining (EDM) involving predicting whether the student’s next attempt will be correct. Many promising ideas have resulted in negligible gains in accuracy, with differences in the thousandths place on RMSE or R2. This talk first briefs common student modeling approaches and how them related to AI, then we explain how well we can expect student modeling approaches to perform at this task. We attempt to place an upper limit on model accuracy by performing a series of cheating experiments. We investigate how well a student model can perform that has: perfect information about a student’s incoming knowledge, the ability to detect the exact moment when a student learns a skill (binary knowledge), and the ability to precisely estimate a student’s level of knowledge (continuous knowledge). We find that binary knowledge model has an AUC of 0.804 on our sample data, relative to a baseline PFA model with a 0.745. If we weaken our cheating model slightly, such that it no longer knows student incoming knowledge but simply assumes students are incorrect on their first attempt, AUC drops to 0.747. Consequently, we argue that many student modeling techniques are relatively close to ceiling performance, and there are probably not large gains in accuracy to be had. In addition, knowledge tracing and performance factors analysis, two popular techniques, correlate with each other at 0.96 indicating few differences between them. We conclude by arguing that there are more useful student modeling tasks are deserving of attention.

Oct 9
Introduced by David Brown
"Computational Creativity: Past, Present and Prosecco"
by Dr. Tony Veale
    For this creative introduction we will be showing a portion of a tutorial video by Tony Veale, a top Computational Creativity researcher, that was delivered at the Autumn School on Computational Creativity 2013. Dr. Veale is from the School of Computer Science and Informatics, University College Dublin. He has been a researcher in the areas of Computational Linguistics, Cognitive Science, Cognitive Linguistics and Artificial Intelligence since 1988, both in industry and in academia. He is the author of Exploding The Creativity Myth: The Computational Foundations of Linguistic Creativity (Bloomsbury Academic, 2012) and a founder member of the international Association for Computational Creativity (ACC).

Oct 16
No Meeting: last day of A term

Oct 23
Break: no meeting

Oct 30
B term: Beckett Conf. Rm.: DCB host: CR coordinator
Qian (Steve) He
"Characterizing the Performance and Behaviors of Runners Using Twitter"
    Running is a popular physical activity that improves physical and mental well-being. Unfortunately, up-to-date information about runners' performance and psychological wellbeing is limited. Many questions remain unanswered, such as how far and how fast runners typically run, their preferred running times and frequencies, how long new runners persist before dropping out, and what factors cause runners to quit. Without hard data, establishing patterns of runner behavior and mitigating challenges they face are difficult. Collecting data manually from large numbers of runners for research studies is costly and time consuming. Emerging Social Networking Services (SNS) and fitness tracking devices make tracking and sharing personal physical activity information easier than before. By monitoring the tweets of a runner group on Twitter (SNS) over a 3-month period, we collected 929,825 messages (tweets), in which runners used Nike+ fitness trackers while running. We found that (1) fitness trackers were most popular in North America (2) one third of runners dropped out after one run (3) Over 95% of runners ran for at least 10 minutes per session (4) less than 2% of runners consistently ran for at least 150 minutes a week, which is the level of physical activity recommended by the CDC (5) 5K was the most popular distance.

Nov 6
Xiangnan Kong
Taming Big Data Variety: From Social Networks to Brain Networks
    Over the past decade, we are experiencing big data challenges in various research domains. The data nowadays involve an increasing number of data types that need to be handled differently from conventional data records, and an increasing number of data sources that need to be fused together. Taming data variety issues is essential to many research fields, such as biomedical research, social computing, neuroscience, business intelligence, etc. The data variety issues are difficult to solve because the data usually have complex structures, involve many different types of information, and multiple data sources. In this talk, I'll briefly introduce the big data landscape and present two projects that help us better understand how to solve data variety issues in different domains. The first project addresses the challenge of integrating multiple data sources in the context of social network research. Specially, I will describe a network alignment method which exploit heterogeneous information to align the user accounts across different social networks. The second project addresses the challenge of analyzing complex data types in the context of brain network research. I will model the functional brain networks as uncertain graphs, and describe a subgraph mining approach to extract important linkage patterns from the uncertain graphs. I'll also introduce future work in this direction and explain some possibilities for upcoming evolutions in big data research.

Nov 13
Artem Gritsenko
Learning Task-Specific Path-Quality Cost Functions From Expert Preferences
    The problem we address in this research is evaluating path plans for a robot operating in complex environments where the cost function for the robot's motion is not readily available. Learning-from-demonstration is often used to recover a cost function from a demonstration of the task. However, providing a demonstration can be difficult for tasks with complex kinematic constraints, especially if the robot being used is highly un-anthropomorphic. Instead, we propose to define task-specific cost functions that represent path quality by learning from an expert's preferences. We tested our method on two simulated car-maintenance tasks with the PR2 robot: removing a tire and extracting an oil filter. We found that learning methods which produce non-linear combinations of the features are better able to capture expert preferences for the tasks than methods which produce linear combinations. This result suggests that the linear combinations used in previous work on this topic may be too simple to capture the preferences of experts for complex tasks.

Nov 20
Seth Adjei
Searching for Optimal Skill Graphs using Data Fitting Techniques
    For some decades now, the Educational Data Mining community has focussed its attention on predicting a student's performance on an item, given data on the students previous performance. Several models including Knowledge Tracing, Performance Factors Analysis, etc have been used to this effect. However there has been little focus on whether prerequisite skill hierarchies can be improved by looking at student log files. A few have proposed models for determining the skills required to solve a given problem, however these approaches have not been linked with the search for optimal prerequisite skill graphs. We therefore report on a number of approaches we have used for improving skill graphs and for identifying close to optimal skill graphs, given an initial skill graph. These approaches include, brute force search with random graph generation, systematically working with an initial skill graph to generate a close to optimal skill graph, and using empirical data to determine the strength of prerequisite skill links. We find some of the approaches useful for the problem at hand and find that additional work is needed to perfect these.

Nov 27
Thanksgiving Break

Dec 4
Andi Dhroso - POSTPONED

Dec 11
Beckett Conf Rm
Anahita Mohseni Kabir
Interactive Hierarchical Task Learning from a Single Demonstration
    We have developed learning and interaction algorithms to support a human teaching hierarchical task models to a robot using a single demonstration in the context of a mixed-initiative interaction with bi-directional communication. In particular, we have identified and implemented two important heuristics for suggesting task groupings based on the data flow between tasks and on the physical structure of the manipulated artifact. We have evaluated our algorithms with users in a simulated environment and shown both that the overall approach is usable and that the grouping suggestions significantly improve the learning and interaction.

Dec 18
(hold for possible thesis presentation)

-- Dec 19 -- Last day of Semester


[Return to the WPI Homepage] [Return to the CS Homepage]

AIRG Coordinator