I create a distinction between the two, by letting the first be a direct
reference to answering the question 'How do brain-like objects achieve their
performance?'. I shall argue that despite years of pleading
modular/architectural studies have not been pursued as much as could have
been done in the last fifteen years of neural networks. Recently solved
problems in the area visual awareness will be discussed and unsolved
problems presented.
|
|
NEURAL REPRESENTATION MODELLING Igor Aleksander |
This will introduce a MECCANO-like technique (NRM - available as shareware
over the Internet) which allows the user to structure and assess modular
structures of digital neural systems of large dimensions. The aim of the
technique is to structure assemblies of recursive modules and do rapid
prototyping of systems. On-line demonstrations will be given. The
educational aspects of this approach will be stressed.
|
MULTIAGENT ROBOTIC SYSTEMS Ronald Arkin |
Research conducted within the Mobile Robot Laboratory at Georgia Tech
has been studying many important issues involving multirobot teams.
We first review an early study regarding the role of
communication in multiagent robotic systems. Results from more recent
areas of multiagent research are then presented (time permitting):
|
BEHAVIOUR-BASED ROBOTICS Ronald Arkin |
Traditional approaches to planning and control using artificial
intelligence techniques for the navigation of mobile robots have generally been based
upon reasoning over abstract models of the world. These models are
either created from a priori knowledge or are derived from sensory
information. Decision-making is based upon this abstracted representation
of reality. It has been shown that there are several pitfalls associated with
this approach, not least of which are the inherently slow response of
these systems and the inaccuracies present due to the lag between the real
world and the abstracted model.
Behavior-based reactive control strategies have been created in response
to the limitations of model-based planning and control techniques. For these
systems, abstract models of the world are avoided in favor of the
immediate utilization of sensory data. In the reactive approach, robot response is
not mediated by a model but is directly invoked from one or more sensory
sources.
After an introductory exposition, we will briefly discuss the motivating
influences for behavior-based robotic systems and their roots in
neuroscience, psychology, and ethology. This discussion is followed by a
presentation of the appropriate role of sensing and representation,
A short survey of exemplar behavior-based systems is then presented with a
critical analysis of each approach.
|
ALGORITHMS FOR COMPUTER VISION Vito Di Gesu' |
The observation of visual forms and patterns has always been pre-eminent
in most of the human activities. Images permeate our life: "stop the car
at red traffic lights ", "select ripe tomatoes, and discard bad ones",
"read the newspaper to update the knowledge" are examples of daily life.
Moreover, image analysis is at the basis of most of the human
activities: astronomers analyze sky maps, radiologists perform diagnoses
by means, for example, of MRI images, and robotic vision is necessary
for autonomous driving.
The advent of digital computers has determined the development of
automatic image analysis systems. The birth of the "modern" computer
vision is related to that of the Cybernetics that can be dated around
1940. In that period the physicist Nobert Wiener and the physician Artur
Rosenblueth promoted, at the Harvard Medical School, meetings between
young researchers, to debate scientific topics. The guideline of those
meetings was the formalization of biological systems (including the
human behavior).
Visual pattern recognition is a process, which develops throughout
several layers of increasing abstraction, corresponding to a set of
iterated transformations. The purpose is to reach a given goal, starting
from an input image scene.
The computation paradigm is conventionally divided in several phases:
from the low level vision processing (examples are filtering, digital
transformations,...) to the Interpretative level vision processing
(examples are semantic description, extraction of physical models).
Different levels of abstraction characterize each of these phases. For
example the low-level vision uses mainly pixel and neighborhood
operators, while the intermediate level vision uses operators that act
on a structured feature space.
The real world is more complex and flexible, all phases interact during
the vision process and a clear distinction between them can't be done.
Therefore their logical sequence has only some relations with the
natural visual process; moreover artificial vision may implements each
visual procedure by using mathematics and physics, regardless to the
neuro-physiologic counter-party.
In this lecture fundamental algorithms in vision are discusses in a
"pragmatic" perspective. At this end, the active vision model of
computation will be considered as a natural evolution of the feedback
mechanism. Links between algorithms and "machine vision" architecture
are also examined. New directions in the design of algorithms for
vision systems are also described.
|
SYMBOLIC, CONCEPTUAL AND SUBSYMBOLIC REPRESENTATIONS Peter Gardenfors |
Within cognitive science, there are currently two dominating approaches to
the problem of representing information. The symbolic approach starts from
the assumption that cognitive systems should be modelled by Turing
machines. The second approach is subsymbolic, mainly instantiated by
connectionism that models cognitive systems by artificial neuron networks.
I will argue that there are aspects of cognitive phenomena for which
neither symbolism nor connectionism offer appropriate modelling tools. I
will advocate a third form of representing information that is based on
using geometrical structures rather than symbols or connections between
neurons. I shall call my way of representing information the conceptual
form since I believe that the essential aspects of concept formation are
best described in this way.
Conceptual representations should not be seen as directly competing with
symbolic or connectionistic representations. Rather, the three approaches
can be seen as three levels of representations of cognition with different
scales of resolution. I will show that the three levels of representation
will motivate different types of computations.
|
CONCEPTUAL SPACES Peter Gardenfors |
A theory of conceptual spaces will be developed as a particular framework
for representing information. I will first present the basic theory and
some of the underlying mathematical notions. A conceptual space is built
up from geometrical structures based on a number of quality dimensions.
Representations in conceptual spaces will be contrasted to those in
symbolic and connectionistic models.
The theory will then be used as a basis for a constructive analysis of
several fundamental notions in cognitive science. Firstly, it will be
argued that the traditional analysis of properties in terms of possible
worlds semantics is misguided and that a much more natural account can be
given with the aid of conceptual spaces. This analysis is then extended to
concepts in general. Some experimental results concerning concept
formation will be presented. In these analyses, the notion of similarity will be in
focus.
Secondly, a general theory for cognitive semantics based on conceptual
spaces is outlined. In contrast to traditional philosophical theories,
this kind of semantics is connected to perception, imagination, memory,
communication, and other cognitive mechanisms.
The problem of induction is an enigma for philosophy of science and it has
turned out to be a problem also for systems within artificial
intelligence.
As a final topic it is argued that the classical riddles of induction can
be circumvented, if inductive reasoning is studied on the conceptual level
of representation instead of on the symbolic level.
|
CONNECTIONIST MODELS FOR DATA STRUCTURES Marco Gori |
Many approaches to learning in connectionist models have two main drawbacks:
First, they cannot process structured information and, second, they learn from tabula
rasa and neglect useful prior knowledge. Whereas algorithms that manipulate symbolic
information are capable of dealing with highly-structured data, adaptive neural networks
are mostly known as learning models for domains in which instances are organized into
static data structures, like records or fixed-size arrays. Structured domains are
characterized by complex patterns which are usually represented as lists, trees,
and graphs of variable sizes and complexity. The ability to recognize and classify
these patterns is fundamental for several applications that use, generate or
manipulate structures (see e.g. applications to molecular biology, classification of
chemical structures, automated reasoning, manipulation of logical terms, software engineering,
recognition of highly-structured patterns, speech and natural language processing).
The purpose of this lecture is that of reviewing significant approaches for overcoming
these limitations. After an introduction to traditional approaches to supervised learning
in neural networks, a unified view of formalisms and tools for dealing with rich data
representations will be presented and early approaches to processing data structures
will be reviewed briefly. Special emphasis will be given on recursive networks,
a natural extension of recurrent networks properly conceived to represent, classify,
and store structured information.
|
SELF-ORGANIZING MAPS I: FUNDAMENTALS Teuvo Kohonen |
The Self-Organizing Map (SOM) is a software tool for the visualization of
high-dimensional data. It converts complex, nonlinear statistical
relationships between high-dimensional data into simple geometric
relationships on a low-dimensional display. As it thereby compresses
information while preserving the most important topological and metric
relationships of the primary data elements on the display, it may also
be thought to produce some kinds of abstraction. These two aspects,
visualization and abstraction, can be utilized in a number of ways in
complex tasks such as process analysis, machine perception, control,
and communication. There exist numerous versions of the SOM,
structural and computational.
|
SELF-ORGANIZING MAPS II: TOPOLOGICAL REPRESENTATION OF MANIFOLDS AND SYMBOL SETS Teuvo Kohonen |
The processing elements, "neurons" of the Self-Organizing Map,
need not be vector-valued models. Each "neuron" can be replaced
by a small network that is able to represent manifolds such as
linear subspaces. Also operator-valued "neurons," equivalent to
dynamical filters, can be used as representations on the map.
The basic SOM usually carries out a clustering in the Euclidean vector
space. Surprisingly, the same vector-space clustering methods
sometimes apply even to entities that are basically symbolic
by their nature. For instance, it is possible to carry out the
clustering of free-text, natural-language documents, if their
contents are described statistically by the usage of different
words in them. Various dimensionality-reduction methods
can be used. Special SOMs for the organization of very large
document collections, called WEBSOM, are described. The largest
WEBSOM constructed so far contains over one million nodes
and is able to map all the electronically available patent
abstracts of the world, seven million in number.
Finally, clustering of completely nonvectorial data such as
symbol strings is possible, too, as long as some distance measure
such as the Levenshtein distance between the data items is definable.
|
MOTOR MAPS AND MOTOR CONTROL MODELS: LEARNING AND PERFORMANCE Pietro G. Morasso |
With the advent of technical means for capturing motion sequences and the pioneering work of Marey and Muybridge, the attempt of describing, modeling and understanding the organization of movement has become a scientific topic. The fact that human movements are part of everyday life paradoxically hides its intrinsic complexity and justifies initial expectations that complete knowledge could be achieved simply by improving the measurement techniques and carrying out a few carefully designed experiments. However, this is not the case, each experiment is frequently the source of more questions than answers and thus the attempt to capture the complexity of purposive action and adaptive behavior, after a century of extensive multidisciplinary research, is far from over. (
continue...)
|
THE SELF-ORGANISATION OF GROUNDED LANGUAGES ON AUTONOMOUS ROBOTS Luc Steels |
The past decade has seen important progress in a
behavior-based, bottom up approach to sensori-motor
intelligence by directly coupling at a subsymbolic level
perception with action. However the problem remains
how we can bridge the gap between the symbolic level
of language, reasoning and problem solving, which is the
domain of "classical AI", and the subsymbolic world
of perception and action. This is the problem that we
have been addressing in recent work, based on the
hypothesis that language might be a key: language
pushes the development of an individual's conceptual
complexity in a co-evolutionary process and through
language a distributed group of autonomous agents may
share conceptualisations of the world.
I will present as a case study the "Talking Heads
Experiment", which we have conducted in the summer
of 1999 as a Turing-test like public
experiment (
http://talking-heads.csl.sony.fr/). The
experiment is based on a set of robot bodies which
have been located in several places in the world
and connected through the Internet. Software agents
can teleport themselves between these robot bodies
and thus experience different realities and engage
in grounded interaction with other agents. The
interaction takes the form of a language game in
which one robot attempts to identify an object
in the environment to the other robot through
verbal means and through pointing gestures.
The language nor the ontology of the robots have been
built in but must be invented and acquired by
the robots autonomously through playing the game. When a game
fails, the robots expand their conceptual repertoires
and/or lexicons. In the experiment, a shared lexicon and
a shared ontology gradually emerges through self-organisation.
The lexicon is grounded in the visual experiences of
the robots.
Humans may interact with the robots
by suggesting words they should use in certain
situations. Thus we have been able to couple the semiotic
dynamics of the artificial language with human natural
language dynamics. Constant ontological and lexical
evolutions are observed as the set of agents is
open (new agents may enter at any time and others may
leave) and the environment is open (new objects may enter
in the environment at any time).
The talk will present the main principles behind
the agent architecture, methods for studying the collective
semiotic dynamics, and results from the experiments.
|
NEURAL MODELLING OF HIGHER ORDER COGNITIVE PROCESSING John Taylor |
Much is now being observed by brain imaging about the global networks
of the brain as they are used to solve different tasks. After a brief
overview of the machines, and how data analysis is performed, the
results being uncovered will be reviewed. The technique of structural
modelling will be described as a method to make clearer the related
networks and their levels of connectivity in the imaging data. The
problem of bridging the gap between the underpinning neural networks
and the observed structural models will then be addressed. The nature
of new paradigms for neural networks to enable higher order cognitive
processing to be modelled, as observed from this and related data,
will be explored. Finally the problem of the representation suporting
consciousness will be discussed.
|
FROM SUBSYMBOLIC TO SYMBOLIC PROCESSING John Taylor |
The problem of obtaining symbolic processing from underlying neural
networks has aroused much controversy. Since the human brain
possesses a solution to this problem, the manner by which this is
achieved will be considered. It will be based on a cartoon model (the
ACTION network) of the frontal lobes (including the basal ganglia and
thalamus). The ability of the ACTION network to learn and regenerate
temporal sequences will be described, and mathematical analysis (using
dynamical systems theory) given to explicate the underlying
mechanisms.
The manner in which this learning ability of the ACTION net gives a
basis for learning of symbol processing, including learning the rules
of syntax, and the ability to manipulate symbols to achieve goals,
will then be explored. The possiblity of giving a neurophysiological
underpinning to the 'deep structures' of Chomsky, and possible
applications of the resulting system to develop a symbol learning and
processing system will be explored to conclude.