|
|
|
|
|
|
|
Abstract
The last 50 years of AI were based on the paradigm of logic-driven search
through symbolic databases. While this has come up with some pretty
smart programs some of the fundamental ways in which the human brain achieves its
intelligence remain undiscovered. This means that there is a dichotomy
between what computer scientists call 'intelligent' and what intellignece is
in living species. In this paper I shall present the three interleaved
forces that are thought to underpin human intelligence. Emergence, once
stripped of its mystical mantle, is a major force in ensuring stability not
only in neural modules but also in architectures of such modules as they
occur in the brain. Evolution can be shown to lead to complex architectures
which are superior to those that engineers design and understand, causing
the direct reverse engineering of the brain to be a precarious business.
Depiction is the opposite of symbolic representation: it relies on the
processing of rich data structures that closely resemble the originating
states of the real world providing the organism whether artificial or real a
sense of conscious control and a sensation of self. Examples of several
current projects on visual awareness and motor control in artificial systems
will be given.
Abstract
The talk introduces and compares models for biologically inspired signal
processing, which are derived from temporally coded neural systems. These
'pulsed' processing models are powerful extensions of the well known classical
neural network paradigms. They have recently become a hot topic in neural
network research, and hardware implementations propose major benefits.
Applications in acoustics and vision will be demonstrated.
Abstract
Although the Human Genome Project will have the entire genome sequence in
2003, determining the role for all genes will remain a challenge. Newer
oligonucleotide microarrays allow the simultaneous quantitative measurement of
the expression of 35,000 unique human RNAs in any tissue. Based on experiments
using this technique, candidate genes linked to pathophyiologic processes are
being found and screened for defects that could predict disease outcomes and
guide optimal therapy. This has successfully been started for diabetes,
leukemia, and other cancers. Linking genomics research to patient care will be
the leading edge of clinical research in the post-genome era.
Abstract
Robotics competitions have become popular and attract people from all walks
of life through out the world irrespective of their ages. Robot soccer makes
heavy demands in all the key areas of robot technology, mechanics, sensors,
communication, and intelligence. The hope of course, is that by discovering how
to get a robot to move with agility, see with acuity, and think perceptively in
the limited context of a soccer game, it will be possible to use the same
techniques to build robots to carry out other more useful tasks. We briefly
discuss the multi-agent scenario pertaining to robot-soccer and its implications
in 21st century. A short discussion on FIRA and related activities will be also
included.
Biography
Professor Normann has moved from conventional electrical engineering to
retinal physiology and into cortical physiology over the past 30 years. His
electrical engineering background and his interests in neurophysiology have
provided unique insights into the studies of the vertebrate visual system. It
has been long appreciated that the parallel processing functions of the nervous
system can best be studied with tools that will allow one to examine firing
patterns of large numbers of neurons or to excite large numbers of neurons via
extrinsic currents. A variety of techniques have evolved which permit this
parallel acquisition of information but many of these techniques have poor
spatial and/or temporal resolution which limit their utility in understanding
how groups of neurons work in concert. Normann and his students and colleagues
have developed a unique microelectrode array which provides unprecedented
spatial and temporal resolution recording of activity from large numbers of
neurons in cerebral cortex and recently, in the peripheral nervous system.
Because these microelectrode arrays are fabricated mainly from silicon, they
have been demonstrated to be highly biocompatible: single unit recordings have
been made from motor cortex in behaving primates for periods exceeding three
years.
Normann and his colleagues have had to develop support systems which enable researchers to utilize these arrays to their fullest capabilities. Implantation of 100 very sharp microelectrodes cannot be achieved with manual techniques. The Normann team has developed a high velocity impulse insertion technique that allows complete insertion of the arrays with little or no cortical trauma. One is next confronted with the problem of having to deal with up to 100 channels of neural information on each array. Normann and his colleagues have developed 100 channel neural signal amplifiers that boost and, optimally filter the signals. The signals are sent to a 100-channel digital signal processing based data acquisition system. This system comprises 100 channels of data, displays on-line and in real time continuously produces raster plots from each of the 100 channels and stores this data in a conventional Pentium-class P.C.
This suite of tools has been used by Normann and his graduate students to study parallel information processing and encoding of visual information by the vertebrate retina, the cat visual cortex and monkey motor cortex. Normann and his colleagues have shown that individual ganglion cells in the turtle are relatively poor classifiers of visual features; however, small groups of ganglion cells allow for the classification of intensity into and color good fidelity. Further, Normann and his colleagues have shown that there are temporal dependencies in this encoding of visual information; response shuffling degrades the classification performance of the groups of ganglion cells. Similar temporal dependencies are seen in ensembles of cells in monkey motor cortex. In collaborative studies done with Dr. John Donoghue at Brown University, we have shown that the volitional intent of a monkey trained to play simple video games can be determined from as few as 15 neurons in motor cortex. Again, shuffled responses degrade the estimation of the monkey's performance suggesting that temporal dependencies in the firing of ensembles of M1 units are involved in the encoding of volitional intent.
These tools and their validation in animal experiments, are leading to human experiments that will be directed at the development of neuroprosthetics systems that will offer new avenues for therapy for those with damaged or diseased parts of their nervous system. Professor Normann plans to continue the development of these electrode arrays and associated technologies, using these arrays in animal experimentation and to begin this new phase of human experimentation.
Abstract
The field of artificial intelligence has dramatically
changed during the past 15 years-or-so. Initially, starting in the fifties,
intelligence was essentially considered to be synonymous with thinking, i.e.
with problem solving, reasoning, and logical deduction. Thinking in turn could
naturally be conceptualized as a sequence of steps, as algorithms, which is
why artificial intelligence was mostly viewed as a sub-discipline of computer
science. During the 1980s, as many people started building robots, the limitations
of viewing intelligence as a computational phenomenon exclusively became obvious:
the idea of mapping sensory stimulation such as camera images onto internal
representations, generating plans of action by logical reasoning, and finally
executing them, simply did not work in the real world. It was clear that a radically
new approach would be required. Rodney Brooks of the MIT Artificial Intelligence
Laboratory suggested that we forget about logic and problem solving, that we
do away with thinking and with what people call high-level cognition and focus
on the interaction with the real world. This interaction is, of course, always
mediated by a body, i.e. the proposal was that intelligence needs to be “embodied”.
What originally seemed nothing more than yet another buzzword turned out to
have profound ramifications and rapidly changed the
research disciplines of artificial intelligence and cognitive science – a new
research field had emerged. It is currently beginning to exert its influence
on psychology, neurobiology, and ethology, as well as engineering.
Embodiment has two main types of implications, physical and information theoretic. The former are concerned with physical forces, inertia, friction, vibrations, and energy dissipation, i.e. anything concerned with the (physical) dynamics of the system, the latter with the relation between sensory signals, motor control and neural substrate. Rather than focusing on the neural substrate only, the focus is now on the complete organism which includes morphology (shape, distribution and physical characteristics of sensors and actuators, limbs, etc.) and materials. Often, problems (e.g. learning problems) that seem intractable if viewed from purely from a computational perspective, turn out to be easy if the embodiment and the interaction with the environment are appropriately taken into account. For example, given a particular task environment, if the morphology and the materials are right, the amount of neural processing required may be dramatically reduced. Moreover, it can be shown that embodiment greatly facilitates the bootstrapping of developmental processes which eventually may lead towards cognition.
In the talk I will review a number of examples illustrating some of the results and novel insights in this field, I will present a set of big challenges for the years to come, and I will outline a few application scenarios.
Abstract
A recent study showed that there were over 400,000 unfilled vacancies in the
IT profession in the USA in large to mid-size companies alone, and that figure
was expected to rise to over a million by the year 2000. While there is certainly
no lack of persons who desire employment in the IT or other E-professions, there
is a critical shortage of qualified professionals. The problem is exacerbated
by mergers and closures in many of the popular technical employment fields,
such as the nuclear and aerospace industries. Many colleges and universities,
while being aware of this situation, are still turning out highly intelligent
graduates, who have been formally instructed in these moribund disciplines,
which are relevant only to an aging and sometimes sentimental professorate,
at the expense of the basic skill sets that all employers require in today's
marketplace. This paper examines an innovative and entrepreneurial approach
to graduate (post baccalaureate) degree programs that have been established
at Penn State Great Valley School of Graduate Professional Studies in Information
Science, Systems Engineering and Software Engineering and in an emerging University
wide degree in Information Sciences and Technology. The paper touches on the
design of efficient degree structures and curricula, and reports some assessment
data collected over the past five years. Degree programs must meet the needs
of both the highly trained (but maybe now redundant) professional and the baccalaureate
novice both of whom are competing for employment in IT and E-Business. To conclude,
the author makes a strong case for colleges and universities to depart from
secure tradition, preferring a sometimes turbulent, but always exciting, new
level of relevancy.
Abstract
Intelligent methods are here defined in the sense of Computational Intelligence, a term that was coined 1992 when the three areas of Fuzzy Technology, Neural Nets, and Evolutionary Computing joined forces to offer solutions to problems, that had so far either no satisfactory or no solutions. Since then these three paradigms have developed considerably and they have
Come up with hybrid methods, where ever they have proven to be complementary to each other. The most important complementarities will be described.
They have first been applied - separately and in combination - to engineering problems, such as, control, process optimisation, analysis, diagnosis and others. Since recently however, applications in various areas of management have proven, that intelligent approaches may have even larger potentials here than in engineering. Successfully completed and ongoing projects will be described and future perspectives discussed.