Our group meets on Thursdays at 11:00 a.m., FL 246, Beckett Conf. Room.
Dates and topics for this semester are as follows:
- Sept 2
- AIRG Organizational Meeting (Coordinator: Joe Beck)
- Sept 9
- Yue Gong
- "How to Construct More Accurate Student Models:
Optimizing Knowledge Tracing and Performance Factor Analysis"
Student modeling is a fundamental concept applicable to a
variety of intelligent tutoring systems (ITS). However, there is not
a lot of practical guidance on how to construct and train such models.
This paper compares two approaches for student modeling, Knowledge
Tracing (KT) and Performance Factors Analysis (PFA), at predicting
individual student trials. We explore the space of design decisions
for each approach and find a set of "best practices" for each. In a
head to head comparison, we find that PFA has considerably higher
predictive accuracy than KT: R2 of 0.18 vs. 0.07 and AUC of 0.76 vs.
0.68 (p<0.001 for both). In addition, we found that PFA's parameter
estimates were more plausible. Our best-performing model was a
variant of PFA that ignored the tutor's transfer model; that is, it
assumed all skills influenced performance on all problems.
- Sept 16
- Dovan Rai
- "Self-disciplined students"
In this study, we are interested to see the impact of
self-discipline on students' knowledge and learning. Self-discipline
can influence both learning rate as well as knowledge accumulation
over time. We used a Knowledge Tracing (KT) model to make inferences
about students' knowledge and learning. Based on a widely used
questionnaire, we measured students' level of self-discipline. When we
analyzed the relation of students' self-discipline with their
knowledge attributes, we found that high self-discipline students had
significantly higher initial knowledge, but there is no consistent
relationship of learning while using the tutor. Moreover, higher
self-discipline students seemed more careful with respect to making
We also found that a quarter of students who were not consistent in
their survey response had significantly lower performance, knowledge
- Sept 23
- Matt Bachmann
The presentation will be about
performing user modeling in an inquiry environment. The purpose of
the talk is to present preliminary results, and to solicit feedback on
how to conduct this MS thesis research.
- Sept 30
- Dave Brown
- "The Curse of Creativity"
Computational design creativity is hard to study, and until fairly
recently it has received very little attention. Mostly the focus has
been on extreme non-routine cases. But there are hard sub-problems and
others ways of moving towards creative systems that are worth
considering. This paper presents three of the alternatives, discussing
one in more depth: i.e., to look at what changes can be made to
routine design systems in order to produce more creative outputs. This
focuses on working "upwards" towards creativity, examining smaller,
ingredient decisions that make a difference to the result. As the
amount of creativity displayed by a design is a judgment made by some
person or group, it should be possible to investigate the degree of
impact of changes to routine design mechanisms. This will contribute
to our understanding of less "extreme" reasoning that leads to
judgments of increased creativity: i.e., the foundation on which other
- Oct 7
- Neil Heffernan and Yutao Wang
Neil will provide an overview of some research done in the
ASSISTment lab for about 20 minutes. Then both will present their most
recent finding. We are trying to see what is the right way to use
hints and attempts in predicting student performance. We have results
we want to share with the group. This work is hot off the shelf.
In fact it is not even on the shelf yet!
- Oct 21
- Andrew Tremblay (first half)
- "Axeawesome: A Return to Green Globs"
Green Globs provided students with "a meaningful and highly
motivating experience with the graphing of equations" and was
celebrated for its novelty, but has since become ignored by
present-day students due to its antiquated form. This project, named
"Axeawesome", tackles the same goal in a modern context, including a
more pleasing aesthetic and online publication with data-mining
capability. This talk will present the current state of Axeawesome, as
well as request advice on how to design the data analysis phase of the
Joseph Beck (second half)
- "Why are we fighting over when to disaggregate? An automated
approach to model construction"
A common discussion point is when to disaggregate data vs.
when to collapse points together. This discussion is another version
of the bias/variance tradeoff in machine learning. This talk will
propose an automated approach to constructing models that
automatically tests various disaggregation possibilities and uses
bootstrapping to decide whether such a disaggregation is sensible.
- Oct 28
- Dan Spitz
- "The Role of Surprise in Design Creativity"
One of the key aspects that people use to judge whether a designed
product is creative is "Novelty". Studies show that this breaks down
into the components "Originality" and "Surprise". In order to produce
a computational system that can produce creative designs it is
conjectured that the system must be able to judge the creativity of
complete or partial designs. Hence, such a system must be able to
judge Novelty. There is research in progress on computational
judgement of originality, and of surprise, but much more of the former.
This talks presents progress on MQP work being done to model surprise
and to build a prototype system that can demonstrate surprise in
- Nov 4
- Jonathan Gibbons
- "Structure optimization for low-end additive manufacturing"
Costs for additive manufacturing processes have dropped well into the
hobbyist range, and are now being used as on-demand manufacturing for
a variety of goods beyond prototypes. This project is exploring
structure optimization methods that incorporate knowledge of the
specific manufacturing constraints and preferences of the fused
filament fabrication process, as used by the sub-$1500 RepRap and
Makerbot machines. Two genetic algorithm methods and an extension of
the solid isotropic material penalization method are being developed,
with the aim of forming the foundation of a practical system to
automatically reduce the material use and fabrication time of objects
produced using these machines.
- Nov 11
- Yue Gong
- "Evaluating web-based resources for educational systems"
Intelligent Tutoring Systems have been shown to have
positive effect on helping student learning. However, almost all types
of the current tutorial content is context-sensitive, thus hard to be
reused in other questions, even for those questions about the same
topic. This talk will propose an idea of using web pages as a new
means to help student learning in an ITS. We will discuss the
framework of how the idea can be applied, as well as relevant research
questions, and how those questions can be answered by using AI
techniques. For example, in order to conduct experiments which are
able to efficiently allocate "trials" for figuring out which web pages
are good, the technique of Reinforcement Learning can be used. Our
goal is to maximize the payoff (determine the usefulness of web
pages), yet use as few trials as possible, while considering the
trade-off between exploration and exploitation. In addition, how to
evaluate the effect of a web page on student learning is another
important problem which could be addressed by applying a variety of
student modeling techniques, such as augmented knowledge tracing
models, learning decomposition. Finally, we want to use machine
learning to extract knowledge (web page features) from the process of
web page evaluations, so that the extracted knowledge can be used in
the future for refining the web page selection. This talk will
discuss these approaches, and solicit feedback for other methods.
- Nov 18
- Jason Zhang
- Meeting cancelled
- Nov 25
- Thanksgiving Break - No Meeting
- Dec 2
- Kaiyu Zhao
- "Structure learning in temporal models"
When learning a model, tradeoff between the capability of
the complexity has always to be made. In this study, a statistical learning
approach is introduced to facilitate the model constructing process,
specifically, the student modeling. Knowledge tracing as one of the well
known models is powerful in interpretation and also tractable in
computation. However, if any extension of KT exists that is able to enrich
the interpretability and avoids the issue of over complicating the real
world, the causalities as well as the changes of the world will be better
captured through the model. A methodology is proposed in finding such models
and preliminary results will be presented together.
- Dec 9
- Yingmei Qi, Shubhendu Trivedi, and Kaiyu Zhao
- "Blind Source Separation in Magnetic Resonance Images"
Blind source separation has received much attention in the past few years
because of its wide range of applications (Neural Networks, Financial Series
Data, Denoising, EEG/MEG data). One important method to do Blind Source
Separation is Independent Component Analysis. And an important application
is in MR Images. A MR Image can be considered to be a mixture as it is
composed of a number of different tissue types such as white matter, gray
matter etc. Source Separation in fMRI is however easier as compared to MRI
as there are enough number of mixtures from which to estimate the hidden
sources. The problem is much harder in MRI as the number of mixtures is just
three (T1, T2 and PD), while the number of sources is around 10, thus being
an under-determined system. Not surpisingly this problem has not recieved
much attention. This project explores ways to solve this problem. Also, ICA
assumes independence of the sources, but it is likely that this assumption
does not hold. We also explore a method to do blind source separation when
some correlation between the sources is assumed. The results are compared
and more ideas on solving this problem presented.
- -- Wed Dec 15 -- Last day of Semester