New Trends in Self-organization and Optimization of Artificial Neural Networks
- Pavel Kordik, Computational Intelligence Group, Department of Computer Science and Engineering, Faculty of Electrical Engineering, Czech Technical University in Prague
- Miroslav Snorek, Computational Intelligence Group, Department of Computer Science and Engineering, Faculty of Electrical Engineering, Czech Technical University in Prague
- Lenka Lhotska, Department of Cybernetics, Faculty of Electrical Engineering, Czech Technical University in Prague
- Martin Macas, Department of Cybernetics, Faculty of Electrical Engineering, Czech Technical University in Prague
- Miroslav Bursa, Department of Cybernetics, Faculty of Electrical Engineering, Czech Technical University in Prague
Time consuming experiments with the identification of optimal topology of artificial neural networks (ANN) are already the past.
ANN users prefer that the topology of the neural network is adjusted automatically to perform optimally on their data set. Optimal structures and proper learning strategies depend heavily on data sets character (dimensionality, complexity of variables relationship, noise, etc.). Therefore, construction of ANN with the optimal structure and best training strategy selection is really hard problem.
Commercial applications typically solve this problem by trial and error ? repetitive generation of random structure and learning strategy while remembering the best-so-far solution.
More efficient approach is to construct the ANN adapted to the problem, which is being solved. Identified complexity of the network should reflect the problem (data set) complexity.
Recently, several techniques based on self-organization of neural network structure and also optimization of its parameters (weights, biases and coefficients) has been introduced. Namely techniques of neural network synthesis using cellular encoding, hybrid nature inspired optimization methods, genetic programming and similar approaches will be central topics of the workshop.
The workshop will be focused namely (but not only) towards neural networks applications in biomedical engineering.
Invited lecture by Kenneth O. Stanley: Abstracting Biological Development to Evolve Large-Scale Artificial Neural Networks
Adaptive Mechanisms of the Perception-Action Cycle
- Vassilis Cutsuridis, Department of Computing Science and Mathematics, University of Stirling, Stirling
FK9 4LA, United Kingdom
- Amir Hussain, Department of Computing Science and Mathematics, University of Stirling, Stirling
FK9 4LA, United Kingdom
- John Taylor, Department of Mathematics, Centre for Neural Networks, King's College London,
London, United Kingdom
The perception-action cycle has been described by many, especially in association
with the notion of affordances of the psychologist JJ Gibson in 1979, being traceable
back to the theoretical biologist JV Uexkull in 1936, but with important developments
by the neuroscientist JM Fuster: "the circular flow of information that takes place
between the organism and its environment in the course of a sensory-guided sequence
of behaviour towards a goal. Each action in the sequence causes certain changes in the
environment that are analyzed bottom-up through the perceptual hierarchy and lead to
the processing of further action, top-down through the executive hierarchy, toward
motor effectors. These cause new changes that are analyzed and lead to new action,
and so on and so forth".
The perception-action cycle and affordances have numerous important processing
features still to be explored. Thus the mirror neurons of Rizzolatti and colleagues
(looked at in the more modern view of internal simulation of the actions of others)
play a role in learning the affordances of objects acted on by another, or of learning
action strategies from the other (using already learnt actions). The full perception-
action cycle itself thus involves these affordance components as well as a hierarchy of
levels in the brain at which this cycle takes place (as the above quotation notes).
Allied to this there is also the development of internal motor models to aid in the
internal simulation of the action-perception cycle, so leading to the possibility of
thinking and more specifically reasoning. As such this whole set of processing
mechanisms is rich with possibilities of extracting, from increased understanding of
how these process work, algorithms able to be inserted into machine systems.
The goal of the workshop is to provide an international, interdisciplinary forum on the
topic of adaptive mechanisms of the perception-action cycle, with the purpose to
advance our understanding of the state-of-the-art on bottom-up and top-down
approaches to artificial cognitive systems development. Presentations and papers on
perception, attention, memory, learning, decision making, reasoning, conflict
resolution, motivation and action are welcome. The manner in which attention is
involved (initially to consciously guide the visual and motor processing and then to let
it run on automatic until error signals bring attention focus back to the source of the
problem and attempt its resolution) will also be considered a highly relevant topic for
the workshop.
The perception-action cycle is an important aspect by which to enter a larger domain
associated with the construction of autonomous machines. The latter require the
perception-action cycle as a basis for development of an embodied system able to
learn (by trial and error or observational learning) how to be increasingly effective in
the given environment of the machine.
Specific Aims
- To bring together a number of leading researchers in the interdisciplinary
fields of cognitive systems, robotics, artificial intelligence and computational
neuroscience;
- To explore current trends on simulation models of autonomous systems with
special focus on its implications for cognitive systems, and through discussion
of implementations of the perception-action cycle, the related aspects of
internal motor models and of internal simulation through mirror-neuron-like
systems, or otherwise;
- To explore the question as to how much knowledge should we take from brain
science in implementing a cognitive system;
- To evaluate levels of investigation, modelling methodologies and current
formalisms regarding the origin and development of cognitive systems.
- To review the main findings and conclusions of previous ICANN "Cognitive
systems" workshops;
- To define a future research agenda for autonomous systems research including
various levels of fielded applications and implementations.
