The Origin of Novelty
31 October 2003
University of Amsterdam
In the framework
of the Cognition programme of the Dutch Research Foundation NWO, the University
of Amsterdam organises a symposium on the origins of novel knowledge and novel
behavior. The symposium is aimed at bringing together ideas on novelty from
Artificial Intelligence/Machine Learning, Cognitive Psychology, Evolutionary
Biology, Philosophy of Science and Philosophy of Language. The workshop is
organised by Maartje Raijmakers (Psychology, Faculty of Social and Behavioural
Sciences), Maarten van Someren (Social Science Informatics, Faculty of Science)
and Jaap Kamps (Information Science, Faculty of Humanities).
The location: Herengracht 182, Amsterdam, Grote
Vergaderzaal. The symposium
starts at 9.15, registration and coffee from 9.00 and ends at 17.30.
Programme:
Now presentations included!
9.15
- 9.30 Introduction
by Peter Molenaar (University of Amsterdam)
9.30
- 10.15 Michael
Thomas (University College
London)
Connectionist approaches to cognitive
variability
10.15
- 11.00 Denis Mareschal (Birkbeck College London)
Novel representations in connectionist networks
11.00
- 11.30 BREAK
11.30
- 12.00 Han
van der Maas (University of Amsterdam)
12.00
- 12.30 Frietson Galis (University
of Leiden)
The evolution of novel functions
12.30
- 13.30 LUNCH
13.30
- 14.15 Peter Flach (Bristol University)
Novelty and discovery - a machine learning approach
14.15
- 15.00 Lorenza Saitta (University
of Piemonte Orientale)
Change of representation and abstraction in knowledge discovery
15.00
- 15.30 BREAK
15.30
- 16.00 Renate Bartsch
(University of Amsterdam)
How can something be new and the same
time be understood? A view from concept formation.
16.00
- 16.30 Theo Kuipers (University
of Groningen)
16.30
- 17.15 Panel discussion
chaired by Johan van Benthem (University of Amsterdam)
DRINKS
Registration
is free but since the number of participants is limited, those who want to participate
must register by sending an email with name, address and affiliation to dr.
Maartje Raijmakers, M.E.J.Raijmakers@uva.nl
before 27 October.
Connectionist approaches to cognitive variability
Dr. Michael Thomas, Birkbeck College, University of
London, UK.
Cognitive development, developmental disorders, and
intelligence all represent challenges to psychology to explain what
neurocomputational parameters determine the conceptual states reachable by the
individual -that is, what thoughts can be thought. Connectionist models have
focused this debate by providing implemented models simulating variability of
each type. In this talk, I evaluate the kinds of computational parameters that
have been proposed to explain highly intelligent vs. stupid representational
systems, to explain conceptual change across cognitive development, and to
explain the atypical representational states found in developmental disorders
such as autism and Specific Language Impairment. The following sorts of
question will be addressed: Do the three forms of variability correspond to
variation across the same computational dimensions or different dimensions? Is
intelligence like have a little bit more 'cognitive development'? And, is there
a special 'golden' computational parameter that makes all representational
systems perform better, thereby explaining the general factor of intelligence?
Novel representations in connectionist networks
Dr. Denis Mareschal
Centre for Brain and Cognitive Development, School of
Psychology
Birkbeck College, University of London
Connectionist networks are ideal for studying the
emergence of representations in dynamic neural computational systems. A close
cousin to the dynamic systems approach, it places greater emphasis on the
content and storage of representations than standard dynamic systems account of
development. In this talk, I will review the different ways in which new
representation can emerge in connectionist networks. This will be illustrated
with simulations in the domains of : (1) infant perceptual development, (2)
children's analogical completion, and (3) conservation of number.
Developmental transitions
Prof. dr. Han van der Maas
Developmental Psychology, FMG, University of Amsterdam
Piaget's constructivist model of stagewise development
has been criticized on theoretical and empirical grounds. Using ideas and techniques
from non-linear dynamical system theory both types of criticism can be
overcome. In this talk the empirical test of a major developmental transition
in proportional reasoning will be used as an example.
The evolution of novel functions
Dr. Frietson Galis
Evolutionary Biology, Leiden University
Evolutionary innovations involve changes at the
genotypic and phenotypic level. For a proper understanding of the evolution of
novelties one, thus, needs to understand the link between genotypic and phenotypic
changes. The mapping of genotype on phenotype is by and large very contorted as
well as discontinuous, rendering the task of investigating the link difficult.
I will discuss both genotypic and phenotypic novelties
and the relationships between them. At the phenotypic level I will focus on
morphological and behavioural novelties. I will discuss the different types of
novelties, the effect of novelties on evolution and the factors that promote or
constrain the evolution of novelties. Emphasis will be on the similarities
between evolutionary novelties in general. An attempt will be made to explore
the relevance for the understanding of the evolution of cognitive novelties.
Novelty and discovery: a machine learning approach
Prof. dr. Peter Flach, University of Bristol, UK
In this talk we describe our own approach to rule
discovery, implemented in the Tertius system. Tertius discovers implications
expressed in first-order logic that are highly confirmed by the data. The
confirmation heuristic is based on a novelty measure that compares the numbers
of supporting and contradicting instances of a rule with the numbers that would
be expected under the null-hypothesis of statistical independence of the rule's
antecedent and consequent. We also discuss how this approach could be used as a
feature-construction step in machine learning.
Change of Representation and Abstraction in Knowledge Discovery
Prof. dr. Lorenza Saitta
University
of Piemonte Orientale, Italy
It is well known that a suitable representation of a
problem can greatly help its solution. The same is true for learning and
discovery: the way in which phenomena and data are perceived/described can let
commonalities and differences emerge, that cannot be "seen" with
alternative representations.
The feature construction and abstraction issues, as
they have been handled in machine learning and knowledge discovery, will be
illustrated and some hints for defining conditions that a "good" representation must
satisfy (given a specific problem) for letting novelty to be grasped will be
discussed.
How something can be new, and nevertheless be understood
Prof. dr. Renate Bartsch, ILLC, UvA
Understanding something is incorporating it into an existing,
previously formed, conceptual system, while preserving the systemıs stability.
On the experiental level this means integrating new
data salva stabilitate into similarity and contiguity sets of previous data.
They have been formed as series of similar and/or contiguous situations or objects under certain
perspectives. For integration into similarity sets, stability means that the
new thing or situation fits into similarity sets representing general concepts
without diminishing the set internal similarity degree. Thatis, the new item
does not add anything conceptually, and in this sense is not really new; it
just fits. For integration into a historical concept, i.e. an individual
concept, an episodic concept , or a concept of an epoch or development,
stability means that the new situation or object fits into the contiguity
structure, i.e. the new situation
is added under coherence of the contiguity between the situations that
form the series of situations that are the partial historical concept.
This unproblematic understanding of an object or
situation is the normal, uninteresting case. (It is still interesting for us on
a meta-level because of its contrast and its relationships to problematic cases
of understanding.) Something really new poses a problem of understanding. It
does not fit, at least not without special cognitive process4es taking place,
so it seems.
What are these processes of problematic understanding?
I shall present the case of metaphorical
understanding, as well as its productive side, namely what is involved in
creating something new by metaphorical projection. This I shall not only treat
for metaphorical language use, but also as a general phenomenon of creating new
situations and objects by relating them in a metaphorical way to series of old,
previously encountered situations and objects. The main step in the creation of
metaphoricity will be perspective change, whereby other similarities (and
contrasts) and other contiguities come into view or are imagined that have not
been operative in the formation of the previous concepts, to which the new
conceptualization can be related as deviating from, but also as a continuation
under a new perspective.
Computational models of scientific discovery, evaluation and revision of concepts, laws and theories
Prof. dr. Theo A.F. Kuipers
Department of Philosophy, University of Groningen
Computational philosophy of science focuses on
computational models for the discovery, evaluation and revision of scientific
concepts, laws and theories. To design such models insights from philosophy of
science are combined with computational models from research in cognitive
science and artificial intelligence, like heuristic search systems and
connectionist networks. Some well-known models (notably models developed by
Simon c.s. and Thagard) and principles will first be indicated, together with
examples from scientific history and current practice. The emphasis will,
however, be put on some challenges for computational modelling of in particular
theory- and domain revision, derived from studies of empirical progress and
truth approximation, and the extent to which Atocha Aliseda and Joke Meheus
have already met them.