2018 Meeting
- When:
- September 7, 8:45am–5:00pm
- Where:
- Davis Center 422 (Jost)
Format
- 10m
- 8m presentation, 2m Q/A
- 20m
- 17m presentation, 3m Q/A
- 30m
- 25m presentation, 5m Q/A
Schedule
08:15–08:45am | Breakfast (bagels, cream cheese, yogurt, granola) |
08:45–09:00am | CEMS Dean Linda Schadler: Welcome |
09:00–10:00am |
[Invited Talk] Hiroki Sayama: Seeking Open-Endedness in Artificial Chemistry Models
Abstract
How to achieve open-endedness in artificial evolutionary systems
has recently become a hot topic in the fields of Artificial Life
and Artificial Intelligence. In this talk, two different
approaches will be presented, both using the Artificial
Chemistry framework. The first approach uses heterogeneous
swarms of self-propelled particles (Swarm Chemistry) that evolve
spontaneously without any explicit fitness evaluation involved.
The second approach uses an even simpler model of random-walking
particles whose higher-order properties are determined by a hash
function (Hash Chemistry). Results obtained and common
challenges arising from these models will be discussed.
|
10:00–10:20am | Break |
10:20–12:00pm |
[10m] Aaron Schwartz (adv. Chris Danforth, Taylor Ricketts, Peter
Dodds): Exposure to urban parks improves affect and reduces
negativity on Twitter
Abstract
Urbanization and the decline of access to nature have coincided
with a rise of mental health problems. Here, we use Twitter to
investigate how mental affect (mood) varies before, during, and
after visits to a large urban park system. We find that affect is
substantially higher during park visits and remains elevated for
several hours following the visit. Visits to Regional Parks,
which are greener and have greater vegetative cover, result in a
greater increase in affect compared to Civic Plazas and Squares.
These results point to the most beneficial types of nature
contact for mental health benefits and can be used by urban
planners and public health officials to improve the well-being of
growing urban populations.
[30m] John Ring (adv. Chris Skalka): (TBD) Abstract
(TBD)
[30m] Kevin Andrew (adv. Asim Zia): Deep Reinforcement Learning for Modeling Human Agents in the Lake Champlain Basin Abstract
Modeling the behavior of human agents in complex systems is a
difficult task which can often be improved upon by introducing
elements of machine learning. As a part of my research, I focus
on using DDQN reinforcement learning to model the behavior of
human and governmental agents in the Lake Champlain Basin to
determine potential behavioral and policy changes in response to
projected climatological and environmental scenarios. This talk
will focus on some of the methods used for integrating
reinforcement learning into agent-based models and preliminary
results from my own work.
[30m] Sam Kriegman (adv. Josh Bongard): Design for a body Abstract
Typically, in AI and robotics, researchers try to realize
intelligent behavior in machines by tuning parameters of a
predefined structure (body plan and/or neural network
architecture) using evolutionary or learning algorithms. However,
these systems tend to be extremely brittle to slight aberrations,
as highlighted by the growing deep learning literature on
adversarial examples. In this talk, I will show robustness can be
achieved by evolving the geometry of soft robots, their control
systems, and how their material properties develop in response to
certain kinds of interoceptive stimulus during their lifetimes.
|
12:00–01:00pm | Lunch banquet (research-active students and faculty only) |
01:00–02:10pm |
[10m] Sage Hahn (adv. Safwan Wshah): Automated Endoleak Detection
through Deep Learning
Abstract
Ruptured Abdominal Aortic Aneurysm's (AAA) represent a serious
surgical risk, and are responsible for around 4500 deaths a year
in the United States alone. The most common treatment method for
AAA is by means of endovascular aneurysm repair a non-invasive
surgery where a stent graft is placed within the aorta, and while
generally successful still carries the risk of complications,
typically in the form an an Endoleak. This work seeks to automate
the detection of Endoleaks within the larger 3D CTA volume
through the application of cascaded convalutional neural
networks. Additional goals include segmentation of the Aneurysm
region, a useful diagnostic tool in tracking Aneurysm growth over
time.
[20m] Jack Felag (adv. Josh Bongard): Modularity in Graphs Abstract
Modularity is becoming increasingly important in design because
of the ability to modify one structure without greatly affecting
other structures. In the first half of the talk, we will explore
the definition of modularity and structures, how to quantify or
search for them, and their applications. In the second half, I
will go over my work regarding developing modular graphs via
evolutionary algorithms.
[20m] Larry Clarfeld (adv. Maggie Eppstein): Assessing Risk from Cascading Blackouts given Correlated Component Failures Abstract
Despite the infrequent occurrence of cascading power failures,
their large sizes and enormous social costs mean that they
contribute substantially to the overall risk to society from
power failures in the grid. Therefore it is important to
accurately capture the risk associated with such events. A
cascading event may be triggered by a small subset of k
components failing simultaneously or in rapid succession. While
most prior work, including our own work into an efficient “random
chemistry” method for risk analysis, has assumed that components
fail independently, this paper proposes a method for deriving
correlated outage probabilities such that pairs of branches that
are proximate in space are more likely to fail together than
distant ones. Combining random chemistry risk analysis with this
approach to correlating outages shows that overall blackout risk
can increase substantially with even small correlations between
component outages. Results of correlated vs. uncorrelated risk
estimates are compared for the 2896 branch Polish Grid test case
under various loads.
[20m] Andrew Becker (adv. James Bagrow): Crowdsourcing Features: Using Crowdsourcing Techniques to Generate Novel Informative Features for Prediction Abstract
One of the key challenges in machine learning is feature
engineering, the creation of the features or predictors that go
into the learning problem. Feature engineering often requires
diverse knowledge and experience, and so crowdsourcing the
ideation of features is a prime candidate for discovering diverse
and useful predictors. We study the problem of supervised
learning when workers simultaneously provide both data and
predictors (new survey questions) online. This process leads to a
distinct pattern of missing values in the data matrix, which
needs new, specialized methods in order to analyze the resulting
data.
|
02:10–02:30pm | Break |
02:30–03:20pm |
[10m] Joshua Powers (adv. Josh Bongard): The effects of morphology
and fitness on catastrophic interference
Abstract
Catastrophic interference occurs when an agent improves in one
training instance but becomes worse in other instances. Many
methods intended to combat interference have been re- ported in
the literature that modulate properties of a neural controller,
such as synaptic plasticity or modularity. Here, we demonstrate
that adjustments to the body of the agent, or the way its
performance is measured, can also reduce catas- trophic
interference without requiring changes to the con- troller.
Additionally, we introduce new metrics to quantify catastrophic
interference. We do not show that our approach outperforms others
on benchmark tests. Instead, by more pre- cisely measuring
interactions between morphology, fitness, and interference, we
demonstrate that embodiment is an im- portant aspect of this
problem. Furthermore, considerations into morphology and fitness
can combine with, rather than compete with, existing methods for
combating catastrophic interference.
[20m] Thayer Alshaabi (adv. Safwan Wshah): Traffic Signs Detection and Geospatial Localization Abstract
This research aims to develop a deep learning-based system to
process a stream of road-images in order to detect and classify
traffic signs, then localize them on a map by estimating their
GPS coordinates. Furthermore, we introduce a new and large-scale
dataset to serve as one of the very few benchmarks for US.
traffic signs recognition (TSR).
[20m] Collin Cappelle (adv. Josh Bongard): Evolving robots which reflect the regularity and modularity present in their environment Abstract
Previous work has shown that robots whose bodies and brains
reflect the modularity and regularity present in task
environments can be evolved in a subset of the total environments
and still retain adequate behavior in the remaining unseen
environments. Future work will look at how to give evolution the
tools necessary to evolve robots with the correct forms of
modularity and regularity without human bias or direction.
|
03:20–03:40pm |
[Faculty Talk] Maggie Eppstein: Predicting the Speed of Evolution,
and Implications for Managing the Evolution of AntiMicrobial
Resistance
Abstract
Large microbial populations generate competing beneficial
mutations. A common assumption has been that evolution acts like
a greedy search algorithm, favoring locally optimal competing
trajectories. However, prior studies have not fully considered
how competition between successive alleles on the same trajectory
impacts adaptation rates along, or likelihood of following,
individual trajectories. We derive a metric that quantifies the
competition between successive alleles along adaptive
trajectories and show that this competition governs the rate of
evolution in simulations on empirically determined fitness
landscapes for proteins involved in drug resistance in two
species of malaria parasite and three species of bacteria. We
show that rates of adaptation vary by orders of magnitude along
different trajectories. In some circumstances, "speedier"
trajectories can out-compete "greedier" trajectories. Our
findings are shown to have important implications for efforts to
manage antimicrobial resistance in real-world settings, and are
relevant to the broader intellectual pursuit of predictive
evolution on complex adaptive fitness landscapes.
|
03:40–04:00pm | Break |
04:00–04:30pm |
[10m] Lindsey Stuntz (adv. Chris Skalka): (TBD)
Abstract
(TBD)
[10m] Ali Javed (adv. Byung Lee): Spatiotemporal trajectories as a new approach for studying concentration-discharge relationships of hydrological events Abstract
Many river water quality constituents such as turbidity,
suspended sediments, and nutrients are predominantly transported
during storm events. The analysis of hydrological systems at
event scales helps to characterize the dynamics and flux of such
constituents. Hydrological events have commonly been analyzed
through study of event concentration-discharge (C-Q) plots and
identification of two-dimensional hysteresis loop patterns in the
C-Q plots. While effective and informative to some extent, this
approach has shortcomings in capturing the temporality of
variables, as it ``collapses'' their values as projected on the
C-Q plane. This study analyzes the categories of hydrological
events using three-dimensional spatiotemporal trajectory plots.
Specifically, computational clustering methods are used to
categorize the trajectories of "moving points" that represent the
measurements from two sensors -- in this study, river discharge
and suspended sediment concentration. This in-progress research
utilizes data from turbidity-based monitoring of suspended
sediment from the Mad River watershed, located in the Lake
Champlain Basin in the northeastern United States. The project
aims toward building classes of spatiotemporal trajectories and
comparison with the existing classes of hysteresis loops that are
currently being used for categorizing storm events.
[10m] Maike Holthuijzen (adv. Brian Beckage): Bias correction for climate predictions from the Weather and Research Forecasting Model: a spatial Bayesian approach Abstract
Over the past few decades, human-induced climate change has
accelerated at an alarming pace. If the risks of climate change
are not well-studied, the impacts to human and environmental
systems could be devastating. it is imperative that we are able
to conduct accurate assessments of climate change impacts at
global, regional, and local scales. However, even the most
advanced climate models produce biased output at resolutions too
coarse for local impact studies. In this study we use a Bayesian
hierarchical spatial model that concomitantly improves the
accuracy and downscales climate model projections. First, we will
fit the Bayesian model to climate projections and make
predictions to actual weather station locations. To downscale the
projections to a finer spatial resolution, we will adjust the
predictions with elevational lapse rates calculated from
historical station data. The bias of the resulting downscaled
projections can be quantified by comparing them to actual station
data. To correct the bias, we will apply the quantile-mapping
technique. Our preliminary results show that, when fitted to
simulated spatial data, our Bayesian spatial model outperformed a
simple linear model in test-set predictive accuracy, especially
when the data exhibited strong spatial autocorrelation.
|
04:30–05:00pm |
[Invited Talk] Hiroki Sayama: How to Survive As an Interdisciplinary Being
Abstract
“Interdisciplinary research” has been a buzzword among
universities and funding agencies for quite some time. However,
being interdisciplinary also brings a lot of practical
challenges. In this talk, I will share some of my own
experiences as an interdisciplinary being in academia,
including how one could manage/avoid/exploit such challenges to
fully entertain this otherwise fun-packed way of life.
|
Last updated September 06, 2018