February 2nd
3:00pm
|
Prof. Bruce Porter [email][web]
Director, Artificial Intelligence Laboratory
Director, Knowledge-Based Systems Research Group [web]
UT Department of Computer Sciences
Dr. Ken Barker
Knowledge-Based Systems Research Group
UT Computer Science Department
Building Large Knowledge Bases from Components
Our experience building the Botany Knowledge Base confirms that knowledge
engineering (i.e. encoding domain knowledge in a computational form)
is the bottleneck in building large knowledge-based systems. The goal
of our research is to develop a new, simpler method for knowledge engineering,
one which constructs large knowledge bases by instantiating and assembling
small, reusable components. This has led us to confront basic issues
of semantic composition, such as the nature of the `building blocks'
and composition operations that enable common-sense reasoning.
|
|
|
Dr. Leslie Pack Kaelbling [email][web]
MIT Artificial Intelligence Laboratory [web]
Why Robbie Can't Learn:
The Difficulty of Learning in Autonomous Agents
In recent years, machine learning methods have enjoyed great success
in a variety of applications. Unfortunately, on-line learning in autonomous
agents has not generally been one of them. Reinforcement-learning methods
that were developed to address problems of learning agents have been
most successful in off-line applications. In this talk, I will briefly
review the basic methods of reinforcement learning, point out some of
their shortcomings, argue that we are expecting too much from such methods,
and speculate about how to build complex, adaptive autonomous agents.
I will back up the speculations with recent results demonstrating that
a small amount of human-provided input can dramatically speed learning
in a real mobile robot.
|
March 9th
3:00pm
|
Paper Discussion: "Theory of Mind for a Humanoid Robot" [pdf]
by Brian Scassellati
MIT Artificial Intelligence Laboratory
If we are to build human-like robots that can interact naturally with
people, our robots must know not only about the properties of objects
but also the properties of animate agents in the world. One of the fundamental
social skills for humans is the attribution of beliefs, goals, and desires
to other people. This set of skills has often been called a theory of
mind. This paper presents the theories of Leslie and Baron-Cohen on
the development of theory of mind in human children and discusses the
potential application of both of these theories to building robots with
similar capabilities. Initial implementation details and basic skills
(such as finding faces and eyes and distinguishing animate from inanimate
stimuli) are introduced. I further speculate on the usefulness of a
robotic implementation in evaluating and comparing these two models.
Note: Brian Scassellati will be giving a job talk on Thursday,
March 29 at 11am in ACES 2.302.
|
March 23rd
3:00pm
|
James Bednar [email][web]
Neural Networks Research Group [web]
UT Department of Computer Sciences
Cara Cashon [email]
Children's Research Laboratory [web]
UT Department of Psychology
About Face: Measuring and Modeling Infants' Face Perception
This will be a two-part presentation on the development of face perception
in infancy. We will focus on one key question: To what extent is face
processing domain-specific, i.e. relying on algorithms and circuitry
specific to faces, and to what extent is it driven by general learning
mechanisms? J. Bednar will review empirical studies from newborns through
2 to 3 months of age, and will present computational modeling results
that show how domain-specific preferences can be integrated seamlessly
into a learning model. C. Cashon will discuss these issues from a behavioral
information-processing perspective and will present data from studies
on infants' (from 4- to 7-months of age) processing of faces. These
results will be contrasted with previous findings with infants' processing
of objects.
|
April 6th
3:00pm
ACES 3.408
|
Paper Discussion:
"Cognitive Multi-character Systems for Interactive Entertainment" [pdf]
by John Funge, Sony Computer Entertainment America &
Steven Shapiro, University of Toronto
Researchers in the field of artificial intelligence (AI) are becoming
increasingly interested in computer games as a vehicle for their research.
From the researcher s point of view this makes sense as many interesting
and challenging AI problems arise quite naturally in the context of
computer games. Of course, the hope is that the relationship is a symbiotic
one so that the incorporation of AI techniques will lead to more interesting
and enjoyable computer games. One question that arises, however, is
how far this process can continue? In particular, what, if any, are
the technical roadblocks to applying new AI research to interactive
entertainment, and what would be the expected benefits? In this paper,
we will therefore take a critical look at some AI techniques on the
horizon of our own current research in developing the software infrastructure
required to view interactive entertainment applications as cognitive
multi-character systems.
For more papers on gaming and AI visit the AAAI
Syposium on Artificial Intelligence and Interactive Entertainment.
|
April 20th
3:00pm
ACES 2.402
|
Prof. Peter Macneilage [web]
[email]
UT Department of Psychology [web]
A Lowly Origins View of Speech
The Chomskyan view of the evolution of speech is that it results from
a genetic mutation that gave us, totally from scratch, an abstract innate
knowledge of sound categories and sound sequencing patterns. Neodarwinian
theory requires descent with modification not a biological "big bang"
for speech. What was available to be modified and what modifications
occurred? The "Frame/Content" theory of evolution says that speech began
when we superimposed an already existing capacity for mandibular oscillation
(jaw opening and closing) on vocal fold vibration (voice) to form syllable
frames - open for vowels, closed for consonants. Beyond this, I will
argue that particular sound patterns found in the simple output of modern
infants, which are also present in languages, might have been in the
earliest language because they involve basic biomechanical properties
of the production system, and self organizational pressures, both of
which were present in the earliest speakers.
|
April 27th
3:00pm
ACES 2.402
|
Prof. Thomas R. Shultz [web]
[email]
Director, Laboratory for Natural and Simulated Cognition [web]
Department of Psychology [web]
McGill University [web]
Recruitment Algorithms in Neural Network Modeling
Artificial neural networks (ANNs) that grow their own internal topology
as well as learn their connection weights show a number of advantages
over static ANNs in terms of learning speed, ability to learn difficult
problems, and cognitive modeling. One of these so-called generative
learning algorithms, cascade-correlation (CC), builds a network topology
by recruiting new hidden units into the network when error can no longer
be reduced. CC has been applied to a variety of problems in cognitive
development, including PiagetŐs conservation task. Such simulations
have clarified a number of longstanding developmental issues about knowledge
representation, representation change, stages, transitions, perceptual
effects, and constructivism. A principal limitation of ANN simulations
is that they fail to use existing knowledge in new learning by always
beginning with random connection weights. A new extension of CC, called
Knowledge-based Cascade-correlation (KBCC), overcomes this limitation
by being able to recruit its own previously learned networks, in competition
with single hidden units. Recruitment of relevant knowledge can significantly
speed learning, and it has potential for building more accurate cognitive
models.
Recommended Reading:
Buckingham, D., & Shultz, T. R. (2000). The developmental course of
distance, time, and velocity concepts: A generative connectionist model.
Journal of Cognition and Development, 1, 305-345. [PDF]
Shultz, T. R., & Rivest, F. (2000). Using knowledge to speed learning:
A comparison of knowledge-based cascade-correlation and multi-task learning.
Proceedings of the Seventeenth International Conference on Machine
Learning (pp. 871-878). San Francisco: Morgan Kaufmann. [PDF]
|
May 4th
3:00pm
ACES Auditorium
|
Harold Henry Chaput [web][email]
UT Department of Computer Sciences [web]
Built to Serve: The Ethics of Engineering a Slave Race
The field of Artificial Intelligence states as its goal the creation
of an intelligent being through means other than reproduction. The purpose
of this endeavor is to get machines to do tasks that can currenly only
be performed by humans. It is not science fiction to say, then, that
AI is working towards building a class of intelligent beings to perform
labor. And yet, outside of science fiction, very little time is spent
considering the moral implications of this work. Are the scientists
and engineers building intelligent agents prepared to deal with the
consequences of their success? There are many reasons offered for dismissing
these concerns, most of which have been used throughout history to justify
human slavery.
But perhaps the most prevalent dismissal is the belief that artificially
intelligent beings are impossible, or at least very far in the future.
However, the accellerating rate of innovation and discovery in computer
science, neuroscience and psychology make the goal of AI seem more possible
and proximal than ever. Moreover, I contend that, given the goal of
AI to create intelligent beings, arguments about its possibility or
timeframe do not absolve scientists of their responsibility. The first
artificially intelligent being that is activated will be the dawn of
a new age in humanity, and how that being is treated will set a precedent
from that day forward. However remote or unlikely that day may be, we
should start preparing for it now.
|
