Results 1 
9 of
9
Coloring Black Boxes: Visualization of Neural Network Decisions
 INT. JOINT CONFERENCE ON NEURAL NETWORKS
, 2003
"... Neural networks are commonly regarded as black boxes performing incomprehensible functions. For classification problems networks provide maps from high dimensional feature space to Kdimensional image space. Images of training vector are projected on polygon vertices, providing visualization of netw ..."
Abstract

Cited by 15 (8 self)
 Add to MetaCart
Neural networks are commonly regarded as black boxes performing incomprehensible functions. For classification problems networks provide maps from high dimensional feature space to Kdimensional image space. Images of training vector are projected on polygon vertices, providing visualization of network function. Such visualization may show the dynamics of learning, allow for comparison of different networks, display training vectors around which potential problems may arise, show differences due to regularization and optimization procedures, investigate stability of network classification under perturbation of original vectors, and place new data sample in relation to training data, allowing for estimation of confidence in classification of a given sample. An illustrative example for the threeclass Wine data is described. The visualization method proposed here is applicable to any black box system that provides continuous outputs.
Visualization of Hidden Node Activity in Neural Networks: I. Visualization Methods
 IN PROCEEDINGS OF THE INTERNATIONAL CONFERENCE ON ARTIFICIAL INTELLIGENCE AND SOFT COMPUTING
, 2004
"... Quality of neural network mappings may be evaluated by visual inspection of hidden and output node activities for the training dataset. This paper discusses how to visualize such multidimensional data, introducing a new projection on a lattice of hypercube nodes. It also discusses what type of infor ..."
Abstract

Cited by 10 (7 self)
 Add to MetaCart
Quality of neural network mappings may be evaluated by visual inspection of hidden and output node activities for the training dataset. This paper discusses how to visualize such multidimensional data, introducing a new projection on a lattice of hypercube nodes. It also discusses what type of information one may expect from visualization of the activity of hidden and output layers. Detailed analysis of the activity of RBF hidden nodes using this type of visualization is presented in the companion paper.
Visualization of large data sets using MDS combined with LVQ.
"... www.phys.uni.torun.pl/kmk Abstract. A common task in data mining is the visualization of multivariate objects using various methods, allowing human observers to perceive subtle interrelations in the dataset. Multidimensional scaling (MDS) is a well known technique used for this purpose, but it due ..."
Abstract

Cited by 1 (1 self)
 Add to MetaCart
(Show Context)
www.phys.uni.torun.pl/kmk Abstract. A common task in data mining is the visualization of multivariate objects using various methods, allowing human observers to perceive subtle interrelations in the dataset. Multidimensional scaling (MDS) is a well known technique used for this purpose, but it due to its computational complexity there are limitations on the number of objects that can be displayed. Combining MDS with a clustering method as Learning Vector Quantization allows to obtain displays of large databases that preserve both high accuracy of clustering methods and good visualization properties. 1
Computational intelligence methods for information understanding and information management
"... Abstract: Information management relies on knowledge acquisition methods for extraction of knowledge from data. Statistical methods traditionally used for data analysis are satisfied with predictions, while understanding of data and extraction of knowledge from data are challenging tasks that have b ..."
Abstract

Cited by 1 (0 self)
 Add to MetaCart
(Show Context)
Abstract: Information management relies on knowledge acquisition methods for extraction of knowledge from data. Statistical methods traditionally used for data analysis are satisfied with predictions, while understanding of data and extraction of knowledge from data are challenging tasks that have been pursued using computational intelligence (CI) methods. Recent advances in applications of CI methods to data understanding are presented, implementation of methods in the GhostMiner data mining package [1] developed in our laboratory described, new directions outlined and challenging open problems posed. To illustrate the advantages of different techniques, a single dataset is exposed to the manysided analysis.
Transformation Distances, Strings and Identification
"... Computational intelligence methods usually work in vector spaces and are not able to deal with objects that have complex structures. Methods based on similarity may be applied in structural domains. Similarity may be defined by minimal cost needed to transform one object into another. The costs of t ..."
Abstract
 Add to MetaCart
Computational intelligence methods usually work in vector spaces and are not able to deal with objects that have complex structures. Methods based on similarity may be applied in structural domains. Similarity may be defined by minimal cost needed to transform one object into another. The costs of the substitution operations that such transformation is composed from may be treated as adaptive parameters. For strings this leads to a generalization of edit (Levenshtein) distance. This distance is computed using dynamic programming method and applied to the problem of identifying DNA gene promoter sequences.
Color Histogram Classification using NMF
"... Introduction Visual recognition of objects is one of the most challenging problems in computer vision and artificial intelligence. Approaches to solve this problem have focused on using several methodologies of which, the most common is Principal Component Analysis (PCA). PCA was initially used to ..."
Abstract
 Add to MetaCart
Introduction Visual recognition of objects is one of the most challenging problems in computer vision and artificial intelligence. Approaches to solve this problem have focused on using several methodologies of which, the most common is Principal Component Analysis (PCA). PCA was initially used to describe face patterns in a lowerdimensional space than the image space [18]. Other approaches have also focused on this technique to overcome frequent computer vision problems such as the recognition of objects taken under a wide range of conditions (several viewpoints and illumination conditions) [12], or dealing with partial occlusions by using robust estimation techniques [1, 5]. However, PCA based techniques suffer from several difficulties. Mainly, an image projection to a PCA based space depends on the precise position of relevant objects, on the intensity and shape of background zones, and on intensity and color of illumination. Since PCA treats its inputs globally, the relevant ob
Topology Representing Network Map A new Tool for Visualization of HighDimensional Data
"... Abstract. In practical data mining problems highdimensional data has to be analyzed. In most of these cases it is very informative to map and visualize the hidden structure of complex data set in a lowdimensional space. The aim of this paper is to propose a new mapping algorithm based both on the ..."
Abstract
 Add to MetaCart
(Show Context)
Abstract. In practical data mining problems highdimensional data has to be analyzed. In most of these cases it is very informative to map and visualize the hidden structure of complex data set in a lowdimensional space. The aim of this paper is to propose a new mapping algorithm based both on the topology and the metric of the data. The utilized Topology Representing Network (TRN) combines neural gas vector quantization and competitive Hebbian learning rule in such a way that the hidden data structure is approximated by a compact graph representation. TRN is able to define a lowdimensional manifold in the highdimensional feature space. In case the existence of a manifold, multidimensional scaling and/or Sammon mapping of the graph distances can be used to form the map of the TRN (TRNMap). The systematic analysis of the algorithms that can be used for data visualization and the numerical examples presented in this paper demonstrate that the resulting map gives a good representation of the topology and the metric of complex data sets, and the component plane representation of TRNMap is useful to explore the hidden relations among the features.