Towards On-line Noninvasive Brain Computer Interface Systems with Multi Commands and Biofeedback


Andrzej Cichocki
 
 
Abstract

Interest in developing an effective and fast communication interface connecting the human brain and a machine has grown rapidly over the past decade. The brain-computer interface (BMI) would allow humans to operate computers, robots, wheelchairs, prostheses, and other devices, using brain signals only. BMI research may someday provide a communication channel for humans. At the same time this neuro-technology is a working tool and novel paradigm in computational neuroscience that contributes to a better understanding of the brain, and a novel independent interface for human-machine communication that offers new options for monitoring and control and lead to innovative modalities of interaction. In this presentation we discuss our recent trends and EEG experiments of on-line (near real-time) EEG brain machine interfaces with several independent commands (from 3 to 12 commands) and combine it with visual, auditory and/or somato-sensory neuro-feedbacks. Our long term objective is to develop save, reliable and user friendly BMI system based on several paradigms: SSVEP (steady-sate visual evoked potentials), P300 Visual Evoked potential, motoric imagery limbs movements (mu rhythms) and also Auditory Steady State Responses (ASSR). For preprocessing stage and features extraction, we will use blind source separation techniques, especially ICA, NMF and multi-way array (dynamic tensor analysis) decomposition and other advanced signal processing and data mining tools. Brief overview of promising paradigms and computational intelligence tools will be given and preliminary results will be discussed.



Data Visualization and Temporal Data Processing Using Self-Organizing Neural Networks


How to learn highly non-separable data


Włodzisław Duch
 
 
Abstract

Linearly separable data is easy to learn. Current state-of-the-art learning algorithms, such as neural networks or support vector machines, are useful in some nonseparable cases but fail in other cases. Even simple problems with non-trivial logic, like the parity problem, cannot be learned with such algorithms. Many problems in bioinformatics and text analysis require complex logic or discovery of (approximate) logical structure in the data. A simple extension of linear separability to k-separability is introduced, allowing to measure the degree to which data is nonseparable and thus the complexity of learning. Empirical tests show that the existing methods fail with relatively low complexity of data. Visualization of learning dynamics in neural networks shows that frequently separability cannot be achieved, but simpler goals for learning may be set. The simplest alternative goal is k-separability, or the projection of data on a line, and segmentation into intervals that contain same-class samples. Difficult Boolean function problems have been learned using novel learning algorithms.



Data Visualization and Temporal Data Processing Using Self-Organizing Neural Networks


Pablo A. Estévez
 
 
Abstract

In the first part of this talk, an overview of techniques for non-linear projection of high-dimensional data based on self-organizing neural networks is presented. These methods aim at making projections that preserve the inter-point distances and/or the neighborhood measured in the original space. An advanced method for nonlinear projection is proposed, which represents the manifold topology by a graph, and utilizes the geodesic distance as a metric. In the second part, an extension of self-organizing neural network models to temporal data processing is presented. A context model based on Gamma memories is introduced. The proposed models for data visualization and temporal data processing are compared with other methods published in the literature using artificial and real-world data sets.



On Applications of Computational Intelligence Methods in Life Sciences


Jarek Meller
 
 
Abstract

Revolutionary advances in molecular biology and genomics stimulate a growing trend to incorporate into a broadly defined biomedical research, various computational approaches to data mining, knowledge representation, pattern recognition, reasoning and prediction, collectively referred to as Computational Intelligence (CI). This talk presents an overview of selected problems and challenges in the fields of computational biology and bioinformatics that are motivated by these advances. Current applications of CI methods in molecular biology, genomics and related disciplines, their limitations and future challenges are discussed. In particular, we focus on applications of techniques for data mining, pattern recognition, and prediction to problems involving biological sequence, structure and system analysis. Examples of such problems, that are specifically discussed here, include sequence variation and genotype-phenotype correlations, micro-array expression data analysis, as well as the prediction of protein structure and function. One of our goals is to illustrate the importance of the domain knowledge and the choice of the right model (representation) for the problem at hand. In a broader context, we also discuss the interplay between biological systems on the one hand, and computing on the other hand.



Agent-based Virtual Organization


Marcin Paprzycki
 
 
Abstract

Recently we observe a surge in new technologies that are expected to change the way we process information and support workers in an organization. Two of them that are very often mentioned as "disruptive technologies" are: ontologies and software agents. In our recent project we attempt at conceptualizing the way in which these two technologies can be combined and utilized in information management within an organization. In the proposed approach a virtual organization is conceptualized in terms of roles to be played by agents, organization structure and information flow are represented in terms of agent-agent and agent-human interactions, while all resources (e.g. workers, brake pads, books, software artifacts, etc.) are ontologically demarcated. Finally, all information processing is semantically-driven. In the presentation current stated of our work will be summarized.




Assessing the Quality of Rules with a New Monotonic Interestingness Measure Z


Salvatore Greco, Roman Słowiński and Izabela Szczęch
 
 
Abstract

The development of effective interestingness measures that help in interpretation and evaluation of the discovered knowledge is an active research area in data mining and machine learning. In this pa- per, we consider a new Bayesian confirmation measure for "if..., then..." rules. We analyze this measure, called Z, with respect to valuable property M of monotonic dependency on the number of objects in the dataset satisfying or not the premise or the conclusion of the rule. The obtained results unveil interesting relationship between Z measure and two other simple and commonly used measures of rule support and anti-support, which leads to e±ciency gains while searching for the best rules.