Date and Location | Speaker | Topic | Reading |
1/20/2012, 12:00-13:00,
ICS colloquium (Muenzinger D430) |
Tal Yarkoni, University of Colorado |
Bridging
the mind/brain gap with text: a novel framework
for large-scale automated synthesis of functional MRI data
The
explosive growth of the human neuroimaging
literature has led to major advances in understanding of human brain
function, but has also made aggregation and synthesis of
neuroimaging findings increasingly difficult. In this talk, I discuss
some of the major challenges neuroimaging researchers
face, and
describe a novel brain mapping framework (Neurosynth) that uses text
mining, meta-analysis and machine learning techniques to help address
some of these challenges. The Neurosynth framework can be used
to
automatically conduct large-scale, high-quality neuroimaging
meta-analyses, address long-standing inferential problems in the
neuroimaging literature (e.g., how to infer cognitive states from
distributed activity patterns), and support 'decoding'
of broad
cognitive states from brain activity in both entire studies and
individual human subjects. I illustrate these applications with
concrete examples from several domains, and introduce a web
interface that provides access to the data and tools (http://neurosynth.org),
before
concluding with a discussion of future directions and
potential
avenues for integration with other tools.
|
Yarkoni et al. |
1/26/2012, 15:30-16:30 CS Colloqiuum (ECCR 265) |
Ma Schwager, Boston University |
Active
Estimation for
Multi-Robot Teams in Hazardous Environments
This
talk will discuss the problem of
deploying
a group of
robots into an
environment for state estimation tasks (e.g. target localization, field
estimation, or mapping), while avoiding hazards at unknown positions
that cause
the robots to fail. Some regions of the environment may be hazardous,
for example, due to fire, severe weather, caustic chemicals, or the
presence of
adversarial agents. A probabilistic model is formulated, under which
recursive
Bayesian filters are used to estimate the state and hazards online. The
robots
move both to avoid hazards and to provide useful sensor information by
following
the gradient of mutual information. Efforts toward overcoming the
challenges of
decentralization and scalability will also be discussed.
Warning:
this will be a technical talk. I've
included it
because those of you interested in modeling will find it interesting
for the particular matehmatical techniques.
|
|
2/3/2012, 12:00-13:30, Muenzinger E214 | Danny Oppenheimer, Princeton | Using metacognitive
disfluency to improve decision quality
and educational outcomeshttp://psych.princeton.edu/psychology/research/oppenheimer/index.php |
|
2/7/2012, 3:30-4:30, Discovery Learning Center 170 | David McDonald, U. Washington |
Social Computational Systems: A Research Agenda for HCC
|
|
2/10/2012, 12:00-13:00, ICS Colloquium | Alice Healy, University of Colorado, Psychology & Neuroscienc |
Specificity and Transfer of Learning Knowledge is often highly specific to the conditions of acquisition, so there is limited transfer of learning from training to testing. A series of studies is reported examining specificity and transfer of learning in three very different tasks, including digit data entry, speeded aiming, and time production. These studies address a variety of theoretical issues, including those involving mental practice, variability of practice, and task integration. Despite these differences across studies, they converge on the conclusion that specificity and transfer of learning are not mutually exclusive. That is, significant specificity can occur even when participants appear to transfer their learning from training to testing. Furthermore, the studies show that the extent of transfer and its direction (i.e., positive or negative) is largely dependent on the definition of transfer employed, the baseline level during training (i.e., start or end of training), and the dependent measure used to assess performance (e.g., initiation time or execution time). |
Healy & Wohldmann |
2/17/2012, 12:00-13:00, ICS Colloquium | Kurt van Lehn, Arizona State University |
Now that Intelligent Tutoring Systems are as effective as human tutors, how can they become even better?
Abstract: It is often said that human tutors are 2 standard deviations more effective than classroom instruction and that Intelligent Tutoring Systems (ITS) are 1 standard deviation more effective. This hypothesis, which inspired many important studies of human tutoring and many efforts to replicate human tutoring with natural language tutoring systems, now seems false. Although research continues, the current best fitting hypothesis is that both human tutors and ITS have the same effect size, namely 0.75 standard deviations above no-tutoring instruction. The first part of the talk will support this claim with a meta-analysis of relevant experiments, illustrated with specific experiments from several labs. However, this finding does not imply that ITS researchers should declare victory and retire. The studies found that both human tutors and ITS were far from perfect. ITS researchers should continue to try to achieve a 2 standard deviation effect size. The second part of the talk discusses three methods with promising preliminary results: (1) using machine learning to improve tutorial decision making; (2) teaching and fading an explicit meta-cognitive strategy; and (3) prompting reflection during problem solving. |
|
2/21/2012,11-12, DLC 170 | Chris Le Dantec, Georgia Tech | Human-computer communication | |
2/22/2012, 11-12, DLC 170 | Tom Yeh, U. Maryland | Human-computer communication | |
2/27/2012, 4-5, Humanities 125 | Zygmunt Frajzyngier, CU Boulder |
Theoretical bases for differential marking of noun
phrases: the proper domain for argument-adjunct distinction
The immediate aim of the study is to provide an
explanation for why certain noun phrases are formally less marked
(‘arguments’) and others aremore marked
(‘adjuncts’) within a clause. Rejecting the
widespread assumption that verbs have an ‘argument
structure’ and that verbs assign grammatical relations, as
well as the absolutist notions of core and peripheral grammatical
relations being determined by verbs, the study provides a
nonaprioristic explanation for the phenomena that led to the emergence
of such notions. The proposed theory explains a hitherto
unexplained phenomenon, viz., why verbs and nouns having the same
referential meanings across languages have different syntactic
properties.
|
|
3/2/2012, 12:00-13:00, ICS Colloquium | Dan Roth, University of Illinois, Computer Science | Learning From
Natural Instructions Machine learning is traditionally formalized as the study of learning concepts and decision functions from labeled examples, thus requiring representations that encode information about the target function’s domain. We are interested in providing a way for a human teacher to interact with an automated learner using natural instructions which communicate relevant domain expertise to the learner without necessarily knowing a thing about the internal representation or the learning program. This talk focuses on the machine learning aspects of this problem. The key challenge is to learn natural language interpretations without being given direct supervision at that level; for example, a plausible feedback could be the success or failure of the performed instruction. We will present research on Constrained Conditional Models (CCMs), a framework that augments probabilistic models with declarative constraints in order to support learning such interpretations. In CCMs we formulate natural language interpretation problems as Integer Linear Programs and learning of the objective functions is done via constraints--driven learning and through global inference that account for the interdependencies among interpretation’s components. In particular, we will focus on new algorithms for training these global models using easy--to--get indirect supervision signals and show the contribution of indirect supervision to other NLP tasks such as Information Extraction, Transliteration and Textual Entailment. |
recent paper |
3/16/2012, 12:00-13:00, ICS Colloquium | Juia Evans, San Diego State University | The impact implicit learning
and development of conceptual knowledge in children with Specific
Language Impairment
It has recently been suggested that SLI is a domain
general deficit in implicit learning. Ullman and colleagues have argued
that the implicit learning impairments in SLI are restricted to
procedural learning impairments and that these procedural learning
deficits impact the acquisition and use of bound morphology and
syntax; leaving the acquisition and use of the mental lexicon
largely intact in children with SLI. In this talk I will present
behavioral, EEG, and aMEG data from our lab that suggests that
implicit learning deficits in SLI may extend beyond procedural learning
to include other aspects of implicit learning as well; and show how
these implicit learning deficits result in a qualitatively
different developmental trajectory of the acquisition and use of
lexical conceptual knowledge for children with SLI as compared to
typically developing children.
|
reading 1 reading 2 |
3/21/2012, 17:00-18:30, Koelbel 330 | Paul Rozin, University of Pennsylvania | The Psychology Of Food | |
3/23/2012, 12:00-13:00, ICS Colloquium | Jordan Boyd-Graber, University of Maryland, iSchool and Institute for Advanced Computer Studies | Making Topic Models more
Human(e) Imagine you need to get the gist of what's going on in a large text dataset such as all tweets that mention Obama, all e-mails sent within a company, or all newspaper articles published by the New York Times in the 1990s. Topic models, which automatically discover the themes which permeate a corpus, are a popular tool for discovering what's being discussed. However, topic models aren't perfect; errors hamper adoption of the model, performance in downstream computational tasks, and human understanding of the data. However, humans can easily diagnose and fix these errors. We present a statistically sound model to incorporate hints and suggestions from humans to iteratively refine topic models to better model large datasets. We also examine how topic models can be used to understand topic control in debates and discussions. We demonstrate a technique that can identify when speakers are "controlling" the topic of a conversation, which can identify events such as when participants in a debate don't answer a question, when pundits steer a conversation toward talking points, or when a moderator exerts her influence on a discourse. |
|
4/6/2012, 12:00-13:00, ICS Colloquium | Nikolaus Correll, Comp Sci, University of Colorado | swarm robotics (Nikolaus is giving a CS colloquium earlier in the semester. This will be a different talk, and aimed more for a cog sci audience.) |
|
4/13/2012,
12:00-13:00, ICS Colloqiuium CANCELLED |
Greg Burns, Emory University |
Neuroimaging of Brain-Culture
Interactions I will present the results of two studies that examine the effects of society and culture on individual brain regions associated with decision making. 1) Sacred values, such as those associated with religious or ethnic identity, underlie many important individual and group decisions in life, and individuals typically resist attempts to trade-off their sacred values in exchange for material benefits. We utilized an experimental paradigm that used integrity as a proxy for sacredness and which paid real money to induce individuals to sell their personal values. Using functional magnetic resonance imaging (fMRI), we found that values that people refused to sell (sacred values) were associated with increased activity in the left temporoparietal junction and ventrolateral prefrontal cortex, regions previously associated with semantic rule retrieval. This suggests that sacred values affect behavior through the retrieval and processing of deontic rules and not through a utilitarian evaluation of costs and benefits. 2) Finally, we use neuroimaging to predict cultural popularity - something that is popular in the broadest sense and appeals to a large number of individuals. We used fMRI to measure the brain responses of a relatively small group of adolescents while listening to songs of largely unknown artists. As a measure of popularity, the sales of these songs were totaled for the three years following scanning, and brain responses were then correlated with these "future" earnings. Although subjective likability of the songs was not predictive of sales, activity within the ventral striatum was significantly correlated with the number of units sold. These results suggest that the neural responses to goods are not only predictive of purchase decisions for those individuals actually scanned, but such responses generalize to the population at large and may be used to predict cultural popularity. |
reading 1 reading 2 |
4/18/2012, 10:30-11:30 Discovery Learning Center (Engineering Center) DLC 170 | Sean Munson, University of Michigan |
Preferences & Nudges
in Sociotechnical Systems
Every day,
millions of people make decisions in digital environments or supported
by software tools. Designers of sociotechnical systems influence the
choices people make, both intentionally and inadvertently, with their
design decisions.
In this
talk, I will discuss my research on individual preferences and systems
designed to nudge people to be their better selves. I will present, in
detail, a study of people's preferences for political opinion
information and efforts to nudge those preferences; this study has
shown the importance of individual differences in selective exposure
theory. I will then give an overview of my current research platforms,
designed to help people live healthier and happier lives, and my
research questions and expected contributions.
|
|
4/26/2012, 15:30-16:30 CS Colloquium |
Andrew McCallum, UMass Amherst | machine learning and
language understanding This talk will be more technical, and aimed at a computer science audience. |
|
4/27/2012, 12:00-13:00, ICS Colloquium | Andrew McCallum, UMass Amherst | machine learning and
language understanding This talk will be different than the 4/26 talk. Students may count each talk separately. |