Hybrid ANNs: Models, Algorithms and Data

Contents of the session

Supervised, unsupervised or reinforcement learning in Artificial Neural Networks (ANNs) has been usually achieved by adjusting the connection weights iteratively using a gradient descent-based optimization algorithm such as BackPropagation. The main problems associated with this kind of algorithms are the necessity of a previously defined architecture for the neural net, their sensitivity to the initial conditions of training, their local character and their restriction to only differentiable surfaces. Much recent research has been done for obtaining neural network algorithms by combining different soft-computing paradigms, resulting in hybrid approaches with the advantages of the different paradigms considered. Bio-inspired optimization algorithms, particle swarm optimization, ensemble training or fuzzy logic together with traditional local search algorithms can provide the basic components for better performing algorithms. Heterogenous datasets can also be considered, as an interesting alternative to improve the accuracy of ANNs. This special session is aimed to cover a wide range of works on hybrid ANNs: combinations of ANNs with other other kind of models (logistic regression, self-organizing maps, support vector machines...), fusions of learning algorithms and heterogenous data structures. We hope that this session can provide a common forum for researchers and practitioners to exchange their ideas and report their latest finding in the area.

This special session will be organized within the "5th International Conference on Hybrid Artificial Intelligence Systems (HAIS'10)". HAIS'10 will be held in San Sebastián, Spain, in June of 2010, and provides an interesting opportunity to present and discuss the latest theoretical advances and real-world applications in the multidisciplinary research field of hybrid intelligent systems.

Topics of the special session include (but are not limited to):

Combination of meta-heuristics (evolutionary algorithms, particle swarm optimization...) and local learning approaches (gradient-based methods, hill climbing...).
Simultaneous optimization of neural network weights and architectures.
Memetic and hybrid algorithms.
Multi-objective optimization of the structure and/or the performance of artificial neural networks.
Ensembles or mixture of classifiers based on evolutionary artificial neural networks.
Hybridization of neural network models using different types of basis functions (kernel or projection functions) or other statistical models.
Learning ANNs using heteroneous datasets, e.g. using privileged information during the training phase.
Support vector machines and kernel methods combined with artificial neural networks.
Evolving transfer functions or learning rules.
Recurrent neural networks optimization, Hopfield nets, probabilist neural nets or self organizing maps optimized using hybrid learning algorithms.
Applications of neural networks to science and real-world problems.
Comparison of these kind of methods and models to standard neural network training techniques: statistical tests, comparative experimental design, critical analysis of comparison methods...

Contributions

Original contributions are sought in the area of the topics covered by this special session. All submissions will be refereed by at least two experts in the field based on originality, significance, quality and clarity. Accepted contributions are to be published in the HAIS'10 Proceedings.

Papers must be prepared according to the LNCS-LNAI style template (see: http://www.springer.de/comp/lncs/authors.html) and must be no more than 8 pages long, including figures and bibliography.

NOTICE: At least one author of each accepted paper must register in order for the paper to be included in the HAIS 2010 Proceedings.

Authors are requested to submit their contributions through the HAIS'10 paper submission page and also email a copy to the chair of this special session (PDF format).