AI 10 to Watch

Published on June 2016 | Categories: Documents | Downloads: 35 | Comments: 0 | Views: 194
of 11
Download PDF   Embed   Report

Comments

Content

T h S e e cF t u i t o u n r e T i ot fl e A I

AI’s

10 to Watch
E

very two years, IEEE Intelligent Systems
acknowledges and celebrates 10 young stars
in the field of AI as “AI’s 10 to Watch.” These accomplished researchers have all completed their
doctoral work in the past five years. Despite being relatively junior in their career, each one has
made impressive research contributions and had
an impact in the literature—and in some cases, in
real-world applications as well.
Nominations in all subfields of AI were sought
from a wide range of senior AI researchers. A short
list of top candidates was voted on by the award
committee, and then the decisions were finalized
with the entire advisory and editorial boards of
IEEE Intelligent Systems. I would like to take this
opportunity to thank two past editors-in-chief of
IEEE Intelligent Systems, Jim Hendler and FeiYue Wang, who served as the co-chairs of the AI’s

86

10 to Watch award committee and did a great job
managing the nomination and selection process.
The group nominated this year was particularly
strong. It has been a struggle to choose the best of the
best. In the end, the top 10 surfaced with unanimous
support from the advisory and editorial boards.
We’re particularly pleased about the diversity of
the winning group. It’s safe to say that everyone
involved in the selection process has been very
proud of these young stars’ contributions, of what AI
as a community can offer, and how bright the future
of AI can be. We’re sure that young AI students and
researchers will find inspiration from these young
stars, and that the AI community will look forward
to their continued excellence and sustained impact.
Congratulations again to our young colleagues
for winning this special recognition!
—Daniel Zeng

1541-1672/13/$31.00 © 2013 IEEE

IEEE INTELLIGENT SYSTEMS

Published by the IEEE Computer Society

IS-28-03-AI's 10 to Watch.indd 86

31/07/13 9:21 PM

AI’S 10 T O WA T C H

Nora Ayanian

Massachusetts Institute of Technology

Nora Ayanian is a postdoctoral associate in the Distributed Robotics Lab at

the Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts Institute of Technology. She will join the University of Southern California as a WiSE Gabilan Assistant Professor of Computer Science in 2013.
Ayanian has a PhD in mechanical engineering from the University of Pennsylvania. She received a graduate fellowship from the National Science Foundation (NSF) in 2005 and won Best Student Paper in the International Conference of Robotics and Automation in 2008. Contact her at [email protected].

Automatic Multirobot
Coordination

M

ultirobot systems have been applied successfully in manufacturing and
warehousing applications, but they require task-specific code that isn’t trans-

ferable to other applications, amenable to dynamic conditions, or accessible to nonexperts.
To realize the full potential of multirobot systems in areas such as transportation, logistics,
agriculture, and disaster response, we must
enable them to autonomously collaborate and
interact in dynamic, resource-constrained environments with humans and other agents.
The critical need for end-to-end solutions for
multirobot autonomy—starting with highlevel specifications all the way to delivering
code for individual robots—built on distributed multirobot planning and control foundations is my work’s underlying theme.
Imagine deploying hundreds of mobile
sensors to monitor an underground pipeline
without so much as a keystroke. Abstractions
let you decouple high-level behaviors from
a distributed system’s complex, low-level
control, so users can specify high-level
actions and behaviors for the team without
worrying about instructions for individual
robots. I’m currently using these abstractions
to create a simplified interface via an iPad app

MAY/JUNE 2013

IS-28-03-AI's 10 to Watch.indd 87

that, with only simple multitouch gestures,
generates code to navigate and control large
teams of robots.
To enable high-level specification, I’m
developing safe, provably correct, and automatically synthesized control policies, which
in turn must address three subproblems of
multiagent coordination: task assignment,
planning, and feedback control. Each of
these subproblems individually presents a
significant challenge, but distilling them
into a single turnkey solution for distributed
multiagent problems is a necessity for userfriendly systems.
It isn’t enough, however, to operate safely
in controlled environments. Autonomous
systems must interact in the dynamic world,
so they must be equipped with flexible
control policies that can adapt and react to
changing conditions. As humans, we use
local information to make critical decisions
without the need for central oversight, so I’m
www.computer.org/intelligent

currently working on algorithms that allow
autonomous, distributed systems to make
decisions based on the wealth of contextual
information that humans use.
Bringing the human in the loop can help
create a data-driven approach to group control, as well as inspire new capabilities for distributed autonomous multiagent systems. At
MIT, we’re analyzing data from experiments
involving large crowds of people, forming
patterns with limited instruction and feedback,
limited to local interaction. A surprising
lesson we’ve learned so far is that distributed
localization is quite difficult for humans!
Within the next few decades, autonomous
multiagent systems will play an integral and
visible role in our daily lives. From transporting
people and packages to monitoring environment and infrastructure to distributing resources such as energy and water, their ubiquity
will present significant usability challenges,
necessitating the development of end-to-end
solutions that start with high-level specifications and deliver specialized code for the
entire system. My research is in developing the
technologies that lower the barrier to entry for
multirobot systems.
87

31/07/13 9:21 PM

Finale Doshi-Velez
Harvard Medical School

Finale Doshi-Velez is an NSF postdoctoral fellow at the Center for Biomedical Informatics at Harvard Medical School. Her research focuses on developing novel probabilistic modeling techniques to create clinical diagnostic tools
and generate scientific research hypotheses in medicine. Doshi-Velez has a
PhD in computer science from MIT. She is a 2007 Marshall Scholar and was
a Trinity Prince of Wales Research Student at the University of Cambridge.
Contact her at [email protected].

AI’S 10 T O WA T C H

Defining Diseases
with Data

C

ombined with the big data revolution in medicine, latent variable models promise to help us redefine disease along lines that ultimately matter for treatment.

Traditional disease definitions are based on expert-defined diagnostic criteria.

When a patient presents with a set of symptoms,
the physician uses these criteria to match the
patient to some “hidden” underlying disease.
However, as more clinical data becomes
available in electronic form, we can now turn
this around and ask, “What sets of hidden
diseases best match the symptoms that we
see in our clinical population?” Data-driven
approaches to defining disease represent a
fundamental shift in medicine to more evidencebased diagnosis and treatment.
Latent variable models in machine learning are generally trained to make good
predictions. Hidden states—in our case,
diseases—are inferred because they provide predictive information about observed
sets of symptoms, rather than being defined
by existing standards of practice. This datadriven approach to defining disease has two key
benefits over traditional approaches: defining
disease subtypes based on predictive criteria lets
us use these subtypes to more accurately make
88

IS-28-03-AI's 10 to Watch.indd 88

predictions based on a patient’s characteristics,
and data-driven disease definitions can drive
new scientific hypotheses about the true
underlying causes of related diseases.
My current work focuses on deriving
data-driven phenotypes for three complex,
heterogeneous diseases—autism spectrum
disorder, type 2 diabetes, and inflammatory
bowel disease. Patients with these diseases
don’t fall neatly into traditional classification
criteria, and our goal is to define a patient’s
subtype based on empirical criteria such
as diagnostic codes and laboratory tests.
For example, our clustering analyses have
revealed subgroups within autism at high risk
for major psychiatric disorders, as well as
those with increased rates of gastrointestinal
disorders. The discovery of the first subgroup
prompted additional research to identify
high-risk patients, while the second led to a
study to identify common genetic causes of
inflammatory bowel disease and autism.
www.computer.org/intelligent

Central to all of these approaches is the
use of latent variables. I use Bayesian nonparametric approaches to flexibly model the
number of latent variables in a dataset, because
they scale the model’s sophistication depending
on the complexity of the observations. More
generally, the language of Bayesian graphical
models lets me integrate information across a
variety of (extremely) noisy and incomplete
clinical databases, expert-curated medical
ontologies, and popular text sources. While
models have been developed for specific clinical
data sources, a core technical component of my
research involves developing novel models and
inference techniques to capture structure from
large, truly heterogeneous data.
My doctoral work in Bayesian nonparametric statistics—in which I showed how
these techniques can produce state-of-the-art
results in general sequential decision-making
problems—and my current position at a medical
school put me in a unique position to drive new
science based on these sophisticated machinelearning techniques. Of course, redefining
disease is only the first step: with better disease
models, we can start the journey toward more
personalized, evidenced-based treatment.
IEEE INTELLIGENT SYSTEMS

31/07/13 9:21 PM

AI’S 10 T O WA T C H

Heng Ji

City University of New York

Heng Ji is an associate professor in the computer science department at
Queens College and a doctoral faculty member in the Departments of Computer Science and Linguistics at the Graduate Center of City University of
New York. In Fall 2013 she is joining Rensselaer Polytechnic Institute as an
associate professor and the Edward G. Hamilton Development Chair in Computer Science. Her research interests focus on natural language processing,
especially on cross-source information extraction and knowledge base population. Ji has a PhD in computer science from New York University. She received a Google Research Award in 2009, an NSF CAREER award in 2010,
and the Sloan Junior Faculty award and IBM Watson Faculty award in 2012.
Contact her at [email protected].

Extracting Information
from Heterogeneous Data

R

ecent years have witnessed a big data boom that includes a wide spectrum of heterogeneous data types, from image, speech, and multimedia

signals to text documents and labels. Much of this information is encoded in
natural language, which makes it accessible to some people—for example, those
who can read that particular language—
but much less amenable to computer processing beyond a simple keyword search.
My chosen research area, cross-source
information extraction (IE) on a massive
scale, aims to create the next generation of
information access in which humans can
communicate with computers in natural
languages beyond keyword search, and computers can discover the accurate, concise,
and trustable information embedded in big
data from heterogeneous sources.
IE works by identifying facts, such as a
person’s publicly accessible job history or
location, merger and acquisition activity
from news coverage, disease outbreaks
from medical reports, and experiment
chains from scientific papers. Traditional
IE techniques pull this information from

MAY/JUNE 2013

IS-28-03-AI's 10 to Watch.indd 89

individual documents in isolation, but users
might need to gather information that’s
scattered among a variety of sources (for
example, in multiple languages, documents,
genres, and data modalities). Complicating
matters, these facts might be redundant,
complementary, incorrect, or ambiguously
worded; the extracted information might
also need to augment an existing knowledge
base, which requires the ability to link
events, entities, and associated relations to
Knowledge Base entries.
In my research, I aim to define several
new extensions to the state-of-the-art IE
paradigm beyond “slot filling,” getting to
the point where we systematically develop
the foundation, methodologies, algorithms,
and implementations needed for more
accurate, coherent, complete, concise, and
most importantly, dynamic and resilient
extraction capabilities.

www.computer.org/intelligent

More specifically, my research aims to
answer the following questions:
r How can we ensure global coherence
and conduct inferences to reduce uncertainty? (My research is developing novel
inference frameworks to identify and resolve morphed and implicit information.)
How can we accurately translate the
extracted facts into another language?
(My research involves information-aware
machine translation.)
How can we adapt methods from one genre
to another, from one domain to another?
(My research combines natural language
processing and social cognitive theories.)
r How will we discover and fuse information from noisy data in multiple data modalities such as text, speech, image, and
video? (My research involves developing new representation and methodology
for multimedia information networks.)
Big data offers an explosive amount of
material to mine, and IE techniques will
help all of us make sense of it.

89

31/07/13 9:21 PM

Brad Knox

Massachusetts Institute of Technology

Brad Knox is a postdoctoral researcher in the MIT Media Lab at the Massa-

chusetts Institute of Technology. His research interests span machine learning,
human-robot interaction, and psychology, especially machine-learning algorithms that learn through human interaction. Knox has a PhD in computer
science from the University of Texas at Austin. He won the best student paper
award at AAMAS in 2010, and his dissertation, “Learning from HumanGenerated Reward,” received the Bert Kay Dissertation Award from his
doctoral department and was runner-up for the Victor Lesser Distinguished
Dissertation Award. Contact him at [email protected].

AI’S 10 T O W A T C H

Learning through
Human Interaction

T

o serve us well, robots and other agents must understand both what we
want and how to achieve it. To this end, my research aims to create robots

that empower humans by interactively learning from them. Conventional robot
learning algorithms are rigid—optimizing a
task objective predefined by a practitioner of
AI—and are often slow to learn. Interactive
learning methods address both of these limitations by enabling technically unskilled end-users to designate correct behavior and communicate their task knowledge.
My research on interactive learning has
focused on algorithms that facilitate teaching
by signals of approval and disapproval from
a live human trainer. Operationalizing these
signals as numeric reward in a reinforcement
learning framework, I ask: Given the reward
that technically unskilled users actually
provide, how should a robot learn from these
signals to behave as desired by the trainer?
In one line of this research, I consider how
to learn exclusively from these signals. To
this end, I developed the Training an Agent
Manually via Evaluative Reinforcement
(TAMER) framework. TAMER is a myopic
(that is, valuing near-term reward while

90

IS-28-03-AI's 10 to Watch.indd 90

ignoring long-term reward) and model-based
reinforcement learning approach that has been
successfully implemented to teach a number of
tasks, including Tetris and interactive navigational behaviors on a physical robot. More
recently, I examined the impact of the chosen
reinforcement learning objective on task
performance, focusing on the rate at which
the robot discounts future reward and whether
the robot experiences separate episodes of
learning. This investigation led to the first
successful training of robots that maximize
human-generated reward non-myopically.
Such non-myopic learning promises to shift the
burden from users to the robot, presenting new
algorithmic challenges.
I also considered how robots can learn from
both human reward and a predefined evaluation
function (that is, a reward function from a
Markov Decision Process), combining TAMER
with more conventional learning. In this setting,
the evaluation function is given authority
www.computer.org/intelligent

to determine correct behavior—the robot’s
performance is judged solely by the evaluation
function’s output—and the trainer’s feedback
provides guidance. Compared to the learning
speed and final performance of reinforcement
learning without human training, this research
produced considerable improvements.
In the near future, I hope to confront some
of the new challenges raised by this work,
including scaling up non-myopic learning to
complex, real-world applications; combining
these algorithms with demonstration-based
teaching methods; and learning from implicit
human communication such as smiling or
giving attention. In my postdoctoral research,
I continue to work on learning from human
reward while also developing a robotic reading
companion for young children, trained by
interactive demonstrations from parents.
Through this research into interactive robot
learning, my goal is to connect robots to our
expertise and desires. With interactive learning,
humans are more than passive beneficiaries of
fully autonomous robots. Instead, they actively
understand and exert control over the behavior
of robots, progressing towards a humancentered artificial intelligence.
IEEE INTELLIGENT SYSTEMS

07/08/13 8:06 PM

AI’S 10 T O WA T C H

Honglak Lee

University of Michigan, Ann Arbor

Honglak Lee is an assistant professor of computer science and engineering at
the University of Michigan, Ann Arbor. His research interests lie in machine
learning, which spans representation learning, unsupervised and semisupervised learning, transfer learning, graphical models, and optimization. Lee has
a PhD in computer science from Stanford. He received best paper awards at
ICML and CEAS, and the Google Faculty Research Award. Lee has served as
an area chair for ICML 2013 and as a guest editor of the IEEE Transactions on
Pattern Analysis and Machine Intelligence special issue on learning deep architectures. Contact him at [email protected].

Learning Representations
from Data

M

achine learning has successfully tackled many problems related to realworld applications in artificial intelligence, but the ultimate performance

of machine learning systems critically relies on the quality of features.
To date, most state-of-the-art features are handcrafted by domain experts who put a great deal
of efforts into domain-specific knowledge. However, hand-crafted features can’t capture highlevel semantics or be adapted to training data,
which often yields suboptimal performance.
Further, in some problem domains, such handdesigned features may not be readily available.
Therefore, the problem of feature construction
is a fundamental challenge in machine learning.
To address this challenge, my research
focuses on developing algorithms for learning
useful feature representations automatically
from data—a concept broadly called representation learning. My research’s central theme
is to develop generative and discriminative
learning models by combining properties
such as distributed representation, sparsity,
hierarchical structures, invariance, and scalability to high dimensionality.
Over the years, I’ve made important contributions to this rapidly growing field. For example,

MAY/JUNE 2013

IS-28-03-AI's 10 to Watch.indd 91

I developed some of the key algorithms for learning sparse representations—specifically, through
one of the fastest sparse coding algorithms currently in use and a sparsity regularizer for
learning restricted Boltzmann machines, deep
belief networks, and autoencoders. I also developed convolutional deep generative learning algorithms that can effectively learn compositional
feature hierarchies from high-dimensional
data such as images, videos, and audio.
More recently, I’ve been working on models
that combine representation learning with
structured priors. Specifically, I’ve demonstrated
how to integrate conditional random fields
(which can enforce local consistencies in
output space) with a Boltzmann machine prior
(which can enforce global consistencies in
output space) for structured output prediction.
This hybrid graphical model produces both
qualitatively and quantitatively superior results
in image segmentation and labeling. In addition,
I’ve demonstrated that nonparametric Bayesian
www.computer.org/intelligent

priors can be incorporated into hierarchical
distributed representations, which allows for
learning mid-level, attribute-like features in a
weakly supervised setting.
To make representation learning more
robust and scalable, some of my recent work
addresses the following questions: How can
we learn from scratch by jointly learning and
selecting relevant features from noisy data?
How can we tease out factors of variations
from data with deep generative models? How
can we learn invariant representations with the
notion of transformations? How can we learn
better features from multimodal data? How
can we incrementally control the capacity
of a feature-learning algorithm from a large
stream of online data? How can we develop
a hyperparameter-free feature-learning algorithm by exploiting theoretical connections
between unsupervised learning models?
Overall, representation learning has shown
promise in many areas, with potentially
transformative impacts on computer vision,
speech/audio recognition, information retrieval,
robotics, and natural language processing. I look
forward to many more technical breakthroughs
and exciting applications in the near future.
91

31/07/13 9:21 PM

Nina Narodytska

University of Toronto, Canada

Nina Narodytska is a postdoctoral research fellow at the University of
Toronto, Canada. She’s also a visiting researcher in the School of Computer
Science and Engineering at the University of New South Wales. Narodytska
has a PhD in computer science from the University of New South Wales and
NICTA. She received an Outstanding Paper Award for AAAI 2011 and an outstanding program committee member award at the Australasian Joint Conference on Artificial Intelligence 2012. Contact her at [email protected].

AI’S 10 T O WA T C H

Optimization, Social
Choice, and Game Theory

S

ocial factors are an increasingly important component in solving realworld optimization problems in logistics, traffic control, online services,

and many other domains. Classical optimization theory is often concerned with
finding the best strategy for a set of agents,
where agents can be, for example, software
programs, companies, or individuals. One
underlying assumption in many optimization
problems is that agents have common goals,
but in practice, that’s not always true. Moreover, competition between agents introduces
a game theoretic component in solving optimization problems, as agents can behave
strategically.
In my research, I focus on building interfaces between social choice and optimization
theory. In particular, together with my
colleagues from the ADT group at NICTA
and the University of New South Wales led
by my ex-supervisor Toby Walsh, I investigate
efficient techniques for solving user
preference-oriented optimization problems
by leveraging ideas from the domains of
optimization, social choice, and game theory.
One interesting problem at the boundary of
optimization and social choice theories is that

92

IS-28-03-AI's 10 to Watch.indd 92

of user preference elicitation. For example,
if a user wants to buy a car, she can easily
answer whether a given car configuration is
acceptable. However, it might be difficult
for her to write a formal model to find valid
configurations. With my collaborators, we
contributed to this problem by investigating
how computationally hard it is to help the user
define a problem by asking a set of simple
questions about preferences. In particular,
we’ve investigated different ways of eliciting
user constraints by asking queries about
partial solutions.
Another interesting research question arises
when performing preference aggregation over
multiple agents: Given a set of individual
preferences, how do we take all these
preferences into account in building a socially
optimal preference that best represents
the interests of all individuals? A natural
and generic mechanism to combine such
preferences is voting. Here, we’re looking at
www.computer.org/intelligent

strategic behavior in a set of agents, such as
misreporting preferences, bribery, and control.
Together with my colleagues, I’m also
working on the resource allocation problem:
Given a set of indivisible resources or items
and a set of agents that have their own
preferences over these items, how do you
efficiently allocate? We might have several
CPUs available in the cloud and a set of users
who want to access these resources. One
general mechanism is to let agents select items
following a picking order, such as you get to
go first and pick the item that you most prefer,
and then I get to pick the item left that I most
prefer, and so on. This mechanism is one of
the oldest and well-known sharing procedures
that many people have used at least once in
their life. We proved that the best mechanism
is to let agents alternate picking items in turns.
This solves a long-standing open question
and, for the first time, provides justification for
the mechanism found on school playgrounds
all over the world. We also investigated
strategic aspects of this problem and designed
an efficient algorithm for constructing a
picking order that makes strategic behavior
unnecessary.
IEEE INTELLIGENT SYSTEMS

31/07/13 9:21 PM

AI’S 10 T O WA T C H

Ariel Procaccia

Carnegie Mellon University

Ariel Procaccia is an assistant professor in the computer science depart-

ment at Carnegie Mellon University. Procaccia has a PhD in computer science from the Hebrew University of Jerusalem, and was subsequently a
postdoc at Microsoft and Harvard. He is a recipient of the Victor Lesser Distinguished Dissertation Award (2009), a Rothschild postdoctoral fellowship
(2009), an inaugural Yahoo Academic Career Enhancement Award (2011),
and a TARK best paper award (2011). He is currently the editor of ACM SIGecom Exchanges and an associate editor of the Journal of AI Research (JAIR)
and Autonomous Agents and Multi-Agent Systems (JAAMAS). Contact him at
[email protected].

AI and Economics:
The Dynamic Duo

T

he interaction between computer science and economics has created a
fast-growing research area that makes up in dynamism and inspiration for

what it lacks in having a proper name. From the computational side, AI plays an
increasingly significant and visible role; the
synergies between AI and economics are
one of the themes that drive my own work.
It’s perhaps not surprising that AI
techniques facilitate economic approaches.
For example, kidney exchanges offer patients
with kidney failure the option to swap willing
but incompatible donors with other patients,
thereby enabling more transplants from living
donors. Although economics offers insights
into the design of efficient kidney exchanges,
challenges arise in optimizing the number of
swaps, especially in light of the uncertainty
caused by patients and donors dynamically
entering and leaving the exchange pool.
We’ve developed practical stochastic
optimization algorithms that leverage AI
parameter-tuning methods to “look into the
future” and refrain from swapping now if it’s
anticipated that the same donors can enable
more swaps later; these results can potentially
save lives.

MAY/JUNE 2013

IS-28-03-AI's 10 to Watch.indd 93

But AI is also contributing to economics
in subtler ways, as researchers are becoming
aware of the role that AI paradigms can play
in reshaping economic theory. For example,
work in social choice theory, which studies
topics such as voting, typically assumes static
preferences. In contrast, our work draws on AI
research on symmetries in Markov decision
processes to construct a model of social choice
with dynamically evolving preferences.
At the same time, the models and tools of
economics are increasingly being applied
to problems in AI and in computer science
more broadly. Applications of fair division
theory are particularly promising. Indeed,
economists have devised techniques to
allocate divisible goods in ways that guarantee formal notions of fairness; some of these
techniques can be adapted to address modern
technological challenges such as the allocation
of multiple computational resources in cluster
computing environments and multiagent
www.computer.org/intelligent

systems. Our theoretical and experimental
work attempts to bridge the gap between
theory and reality by tackling dynamic settings
in which users can arrive and depart, and
brings us one step closer to creating fair and
practical resource allocation algorithms.
I’m equally excited about applying
social choice theory to human computation
systems, which combine human and machine
intelligence by employing humans to solve
problems that are difficult for computers.
These systems often use voting methods
to aggregate opinions, but in a naïve way.
We’ve established that a principled approach
to voting, based on the literature on voting
rules as maximum likelihood estimators, can
significantly increase the efficiency of human
computation systems.
In light of this lively interaction between AI
and economics—and the special role played
by dynamic environments—I like to think
of AI and economics as the “dynamic duo.”
In the tradition of another famous dynamic
duo—Batman and Robin, not the Korean
hip hop duo—the partnership between the
two fields promises both groundbreaking
technology and societal impact.
93

07/08/13 8:06 PM

Stefanie Tellex
Brown University

Stefanie Tellex will join the computer science department at Brown

University as an assistant professor in September 2013. She is currently a
research scientist at the MIT Computer Science and Artificial Intelligence
Laboratory. Her research interests include probabilistic graphical models,
human-robot interaction, and grounded language understanding. Tellex has a
PhD from the  MIT Media Lab for her work on models for the meanings of
spatial prepositions and motion verbs. Contact her at [email protected].

AI’S 10 T O WA T C H

Talking to Robots

I

n the home, in the factory, and in the field, robots have been deployed for
tasks such as vacuuming, assembling cars, and disarming explosives. As

robots become more powerful and more autonomous, it’s crucial to develop
ways for people to communicate with
them. Natural language is an intuitive and
flexible way of enabling this type of communication. The aim of my research program is to construct robots that use language to seamlessly meet people’s needs.
To understand language, a robot must
be able to map between words in language
and aspects of the external world. At MIT,
my collaborators and I developed the
Generalized Grounding Graph framework
to address this problem. Our framework
creates a probabilistic model according
to the structure of language, defined by
cognitive semantics. We train the model
using large datasets collected from human
annotators via crowdsourcing. By using
data collected from many different people,
our system learns robust models for the
meanings of a wide variety of words.

94

IS-28-03-AI's 10 to Watch.indd 94

Robots must establish common ground
with their human partners to accurately
follow instructions. One way to establish
common ground is for a person to tell
the robot about the world, in the person’s
own terms. I’m developing interfaces to
enable a robot to understand a person’s
descriptions of the world and integrate
them with its own representations. The
resulting semantic map enables the robot
to more accurately follow directions
because its mental model of the world
more closely matches the person’s.
However, no matter how much training
data a robot has, there will always be
failures to understand. To address this
problem, I’m developing ways for robots to
recover from failures by applying the same
strategy a person does: asking a question.
Methods based on information-theoretic

www.computer.org/intelligent

human-robot dialog enable a robot to use
ordinary language to explain what it needs
to an untrained person. The human provides
help that enables the robot to recover
from its failure and continue operating
autonomously.
Language-based interfaces will make
robots accessible and intuitive for a
wide variety of applications. Household
robots will engage in tasks in the home
such as cooking and cleaning. A robot
that understands human language will
enable untrained users to express complex
requirements, ranging from what to
make for dinner to where to put away the
socks. In the factory floor, or in search
and rescue tasks, humans will supervise
heterogeneous teams of robots to assemble
products or search for survivors after
an  explosive event. Using language,
people can quickly deploy a robot team
where it’s  most needed, and robots can
communicate with the human supervisor
about what they’ve discovered and where
they need help.

IEEE INTELLIGENT SYSTEMS

07/08/13 8:07 PM

AI’S 10 T O WA T C H

Jun Zhu

Tsinghua University

Jun Zhu is an associate professor in the Department of Computer Science

and Technology at Tsinghua University. His research focuses on developing
machine-learning methods to understand complex scientific and engineering data. Jun has a PhD in computer science from Tsinghua University. He’s
an active member of the research community, serving as an area chair for
Advances in Neural Information Processing Systems 2013 and a local chair
for International Conference on Machine Learning 2014. His dissertation received the China Computer Federation Distinguished Dissertation Award,
which recognizes the best PhD in China in computer science. Contact him at
[email protected].

Bayesian AI

T

he world is an uncertain place because of physical randomness, incomplete knowledge, ambiguities, and contradictions. Drawing inference

from noisy or ambiguous data is an important part of intelligent systems, where
Bayesian theory serves as a principled
framework of combining prior knowledge
and empirical evidence. The past 20 years
have seen tremendous progress in developing both Bayesian and nonparametric Bayesian methods for resolving model
complexity and adapting to stochastic and
changing environments with data-driven
learning algorithms. However, conventional Bayesian inference is facing great
challenges in dealing with large-scale
complex data, arising from unstructured,
noisy, and dynamic environments such as
the Web, which records massive digital
traces of human activities.
To address these challenges, my
research consists of developing Bayesian
inference methods and scalable algorithms
to address important problems in scientific
and engineering domains. Through my
work, I’ve developed regularized Bayesian
inference, or RegBayes, a computational
mechanism allowing Bayesian inference to

MAY/JUNE 2013

IS-28-03-AI's 10 to Watch.indd 95

directly control the properties of posterior
distributions by imposing posterior
regularization, which provides a significant
source of extra flexibility to incorporate
domain knowledge in rich forms such as
those represented as a logical knowledge
base and those derived from some learning
principle. RegBayes has rich connections
with information theory, optimization
theory, and statistical learning theory.
When the posterior regularization is
derived from the discriminative maxmargin principle, RegBayes sets up a
bridge between Bayesian nonparametrics
and max-margin learning, two important
subfields in machine learning that have
taken largely disjoint paths over the
past 20 years. In addition, I address the
fundamental computational challenges
of scaling up Bayesian methods to huge
and high-dimensional data. I developed
highly scalable inference algorithms for
RegBayes by exploring problem structures

www.computer.org/intelligent

and utilizing recent advances in Markov
chain Monte Carlo methods and variational
methods.
I’ve worked with social scientists,
computer vision researchers, and biologists to develop hierarchical Bayesian
models to understand how social links
are created and how to predict new links;
how natural scene images are composed
with objects, and how to categorize natural
scenes at a near-human level; and how to
incorporate biological domain knowledge
to understand the relationship between
genomic variations among population
and complex diseases. Answers to such
questions are vital to a range of important
applications for public good. My long-term
research goal is to develop AI systems that
can effectively incorporate rich domain
knowledge, cope with various sources
of uncertainty, discover latent structures
of complex data, and adapt to dynamic
environments. To achieve this goal, it’s
important to perform interdisciplinary
research—to that end, I’m collaborating
with neural scientists on developing
Bayesian AI algorithms with strong neural
and biological evidence.

95

07/08/13 8:07 PM

Aviv Zohar

Hebrew University of Jerusalem

Aviv Zohar is a senior lecturer at the School of Engineering and Computer

Science in the Hebrew University of Jerusalem and he is a Golda Meir Fellow.
His postdoctoral work was at Microsoft Research. Zohar has a PhD in computer science from the Hebrew University. His other honors include an award
of excellence from the Israeli Knesset and the committee of university heads,
a Leibniz scholarship during his PhD studies, and a scholarship from the Wolf
Foundation during his MSc. Contact him at [email protected].

AI’S 10 T O WA T C H

Incentives in
Multiagent Systems

D

esigners of multiagent systems set rules by defining their underlying
protocols that define both the basic language of interaction and the behav-

ior that agents should adopt. For example, TCP defines the format of messages
that are sent but also asks communicating parties to lower the rate of data transmission if too many packets are lost (this
is done to keep the network uncongested).
Due to the system’s distributed nature
and each participant’s autonomy, there’s
no way to strictly enforce such behavior.
Agents that gain from deviating from the
prescribed behavior can do so, and the system behaves quite differently than initially
expected. In the case of TCP congestion
control, deviating from the protocol can
cause a congestion collapse that drastically
slows down all traffic passing through
the network.
My research is aimed at analyzing
the incentives of agents in such systems
using tools from AI, game theory, and
economics. I seek to create protocols
in which the recommended behavior is

96

IS-28-03-AI's 10 to Watch.indd 96

also the best course of action for each
participant without compromising other
properties such as system efficiency or
robustness.
Along with my various collaborators,
I’ve explored a wide range of computational systems including core Internet
communication protocols such as Border
Gateway Protocol (BGP) and TCP, where
communicating parties might attempt
to gain more bandwidth or a more
desirable path through the network. These
foundational protocols, while imperfect,
have interesting incentive structures that
help explain their adoption.
In other systems such as peer-to-peer
file sharing (where participants typically
lack the incentives to upload files to
others), incentive issues hinder wider
adoption. My work has focused on the

www.computer.org/intelligent

partial improvements of current protocols
such as BitTorrent and on exploring the
market-like ad hoc solutions that filesharing communities have adopted.
Finally, novel distributed open systems
such as the crypto-currency Bitcoin
continue to emerge, bringing with them
new challenges. One of Bitcoin’s main
strengths is in its incentives for nodes that
authorize transactions. The transaction
fees awarded to these nodes have attracted
many to join Bitcoin’s network and to
invest their computing resources in
securing it. On the other hand, competition
for these very fees, which is expected to
increase in the future, could cause profitmaximizing nodes to behave in ways that
damage the system. I continue to work on
ways to improve the protocol before such
problems are encountered. The Internet’s
rise and the prevalence of computing
devices promise that innovative and
exciting multiagent systems will continue
to appear and that multiagent systems
research will continue to flourish.

IEEE INTELLIGENT SYSTEMS

31/07/13 9:21 PM

Sponsor Documents

Or use your account on DocShare.tips

Hide

Forgot your password?

Or register your new account on DocShare.tips

Hide

Lost your password? Please enter your email address. You will receive a link to create a new password.

Back to log-in

Close