poster session (1)
Tuesday, 2007-09-04 - from 15:00 to 16:00
Application of SOM in a health evaluation system
|Presenting Author: P.K. Kihato|
|Authors: Heizo Tokutaka, Yoshio Maniwa, P.K. Kihato, Kikuo Fujimura, Masaaki Ohkita|
A health evaluation system was constructed which visualizes the living habits and health state from a person's checkup list by using the feature of SOM that multi-dimensional data can be mapped onto a two-dimensional surface. Here, three examples cases are reported. A change to the health region of the map by taking medication was visualized by the SOM from the conventional numerical expression. Also, the specific sick record converges towards the sick region of the map when the disease progresses. However, it was shown and visualized for the sick record not to converge in the case of the metastasis of a cancer even if for the same examinee, the cancer has progressed. Finally, for the display of the health point mark, and the display of the sick record, the spherical surface SOM, is demonstrated to be suited in the visualization.
GalSOM - Colour-Based Image Browsing and Retrieval with Tree-Structured
|Presenting Author: Philip Prentis|
|Authors: Philip Prentis|
This paper describes an image browsing and retrieval application called GalSOM. Bitmap images are described by their colour histograms and sorted using an improved variant of the tree-structured self-organising map (TS-SOM) algorithm. The advantages of using such a system are discussed in detail, and their application to the problem of image theft detection is proposed.
Functional SOM for variable-length signal windows
|Presenting Author: Arnaud De Decker|
|Authors: Arnaud De Decker, Gael de Lannoy, Michel Verleysen|
Functional data, often sampled at high frequency, lead to high-dimensional vectors. The curse of dimensionality makes the latter difficult to handle with standard data analysis tools. Functional data analysis tools take profit of the functional nature of data by projecting them on a smooth basis. This paper shows how to extend functional Self-Organizing Maps (SOM) to signal windows having different lengths. This technique may be applied for example on signal sampled regularly, but for which the duration of each signal is varying; an example concerns electrocardiography (ECG), where the signal is usually cut according to the variable period between two heart beats.
Self-Organisation of Neural Topologies by Evolutionary Reinforcement Learning
|Presenting Author: Nils T Siebel|
|Authors: Nils T Siebel, Jochen Krause, Gerald Sommer|
In this article we present EANT, a method that creates neural networks (NNs) by evolutionary reinforcement learning. The structure of NNs is developed using mutation operators, starting from a minimal structure. Their parameters are optimised using CMA-ES. EANT can create NNs that are very specialised; they achieve a very good performance while being relatively small. This can be seen in experiments where our method competes with a different one, called NEAT, to create networks that control a robot in a visual servoing scenario.
Variable-Density Self-Organizing Map for Incremental Learning
|Presenting Author: Atsushi Shimada|
|Authors: Atsushi Shimada, Rin-ichiro Taniguchi|
We propose a new incremental learning method of Self-Organizing Map. Basically, there are three problems in the incremental learning of Self-Organizing Map: 1. depletion of neurons, 2. oblivion of training data previously given, 3. destruction of topological relationship among training samples. Weight-fixed neurons and weight-quasi-fixed neurons are very effective for the second problem. However the other problems still remain. Therefore, we improve the incremental learning method with weight-fixed neurons and weight-quasi-fixed neurons. We solve the problems by introducing a mechanism to increase the number of neurons effectively in the incremental learning process.
Mapping of the Genome Sequence Using Two-stage Self Organizing Maps
|Presenting Author: Hiroshi Dozono|
|Authors: Hiroshi Dozono, Tekeshi Takahashi|
In this paper, we introduce an algorithm of Self-Organizing Maps(SOM) which can map the genome sequence continuously on the map. The DNA sequences are considered to have the special features depending on the regions where the sequences are taken from or the gene functions of the proteins which are translated from the sequences. If the hidden features of the DNA sequences are extracted from the DNA sequences, they can be used for predicting the regions or the functions of the sequences. In this paper, we propose the algorithms using two stage SOM which organizes the sequences of the specific length at the first stage and organizes the set of sequences at the 2nd stage This algorithm can map the genome sequences on the map at each stage depending on the features of the sequences.
We made some analyses of the genome sequences concerning the functions, species and secondary structure of the sequences.
An Energy Function-Based Optimization of Matching Parameters and Reference Vectors in SOR Network
|Presenting Author: Hideaki Misawa|
|Authors: Hideaki Misawa, Takeshi Yamakawa|
In this paper we propose an energy function-based optimization method in order to improve the approximation ability of the self-organizing relationship (SOR) network. In the execution mode, the SOR network can be used as a fuzzy inference engine. The output of the SOR network is calculated by using the reference vectors and matching parameters. The matching parameters, which correspond to the standard deviation of the Gaussian membership function used in fuzzy inference, are only defined in the execution mode. However, the issue of the optimization of the matching parameters has not yet been treated in previous works. To optimize the matching parameters, we introduce an energy function to the SOR network. The energy function can be used to tune not only the matching parameters but also the reference vectors with a gradient descent method. The proposed method is applied to a function approximation problem and the improvement of the approximation ability is confirmed.
Transform Learning - Registration of medical images using self organization
|Presenting Author: Dietlind Zühlke|
|Authors: Dietlind Zühlke|
A network model is introduced that allows a multimodal registration of two images. It can be used for a image-model or a model-model registration. The application of the network to registering tomographic to 3D ultrasonic data is introduced. Results on artificial and real ultrasound image data sets are discussed.
Composition of Self Organizing Maps for Adaptive Mesh Construction on Complex-shaped Domains
|Presenting Author: Olga Nechaeva|
|Authors: Olga Nechaeva|
In this paper, an important application of Self-Organizing Maps (SOM) to construction of adaptive meshes is considered. It is shown that application of the basic SOM model leads to a number of problems like inaccurate fitting the border of a physical domain, mesh self-crossings, etc. The composite SOM model is proposed which is based on the composition of a number of SOM models interacting in a special way and self-organizing over their own set of input data. A core of the composite SOM model is the colored SOM model with nonadjustable neurons which provides us a technique to control the neuron weights adjustment taking into account the fixed ones and the general layout of the mesh. As a result, the composite SOM model allows us to approximate an arbitrary complex physical domains with well topology preservation.
Path finding on a spherical SOM using the distance transform and floodplain analysis
|Presenting Author: Masahiro Takatsuka|
|Authors: Michael Bui, Masahiro Takatsuka|
Data visualization has become an important tool for analyzing very complex data. In particular, spatial visualization enables users to view data in a intuitive manner. It has typically been used to externalize clusters and their relationships which exist in highly complex multidimensional data. We envisage that not only cluster formation and relationships but also other types of information, such as temporal changes of datum, can be extracted through the spatialization.
In this paper, we investigate an application of trajectory/path analysis carried out using a Self-Organizing Map as a spatialization method. We propose an application of distance transformations to the Geodesic Self-Organizing Map. This new approach allows a user to visually inspect the trajectory of multidimensional knowledge pieces on a two-dimensional space. The trajectories discovered through this approach are essentially the shortest paths between two points on the Self-Organizing Map. However, those paths might go outside of the input dataspace due to the connectivity of neurons imposed by the grid structure. We also present a method to find the shortest path, which falls within the input dataspace using simple floodplain analysis.