WSOM+ 2019: 13TH INTERNATIONAL WORKSHOP ON SELF-ORGANIZING MAPS AND LEARNING VECTOR QUANTIZATION, CLUSTERING AND DATA VISUALIZATION
PROGRAM FOR WEDNESDAY, JUNE 26TH
Days:
next day
all days

View: session overviewtalk overview

08:30-09:00

Registration

09:00-10:15

Welcome speech and invited talk (Alessandro Sperduti, Italy)

10:15-11:05 Session 1

LVQ: Practical Applications

10:15
LVQ-type Classifiers for Condition Monitoring of Induction Motors: A Performance Comparison

ABSTRACT. In this paper, we introduce a design methodology for prototype-based classifiers, more specifically the well-known LVQ family, aiming at improving their accuracy in fault detection/classification tasks. A laboratory testbed is constructed to generate the datasets which are comprised of short-circuit faults of different impedance levels, in addition to samples of the normal functioning of the motor. The generated data samples are difficult to classify as normal or faulty ones, especially if the faults are of high impedance (usually misinterpreted as non-faulty samples). Aiming at reducing misclassification, we use K-means and cluster validation techniques for finding an adequate number of labeled prototypes and their correct initialization for the efficient design of LVQ classifiers. By means of comprehensive computer simulations, we compare the performances of several LVQ classifiers in the aforementioned engineering application, showing that the proposed methodology eventually leads to high classification rates.

10:40
Searching for the Origins of Life -- Detecting RNA Life Signatures Using Learning Vector Quantization

ABSTRACT. The most plausible hypothesis for explaining the origins of life on earth is the RNA world hypothesis supported by a growing number of research results from various scientific areas. Frequently, the existence of a hypothetical species on earth is supposed, with a base RNA sequence probably dissimilar from any known genomes today. It is hard to distinguish hypothetical sequences obtained by computer simulations from biological sequences and, hence, to decide which characteristics provide biological functionality. In the present consideration biological sequences obtained from RNA-viruses are compared with computationally generated sequences (artificial life probes). The task is to discriminate the samples regarding their origin, biological or artificial. We used the learning vector quantization (LVQ) model as the respective classifier, which has only weak requirements regarding the underlying dissimilarity measure. This gives the opportunity to investigate several dissimilarity measures regarding their discriminating behavior for this task. Particularly, we consider information theoretic dissimilarities like the normalized compression distance (NCD) and divergences based on bag-of-word (BoW) vectors generated on the base of nucleotide-codons. Additionally, the geodesic path distance is applied taking an unary coding of sequences for a representation in the underlying Grassmann-manifold. Both, BoW and GPD allow continuous updates of prototypes in the feature space and in the Grassmann-manifold, respectively, whereas NCD restricts the application of LVQ methods to median variants.

11:05-11:30

Coffee Break

11:30-13:10 Session 2

SOM: Theory and Methods

11:30
Look and feel what and how recurrent self-organizing maps learn

ABSTRACT. This paper introduces representations and measurements for revealing the inner self-organization that occurs in a 1D recurrent self-organizing map. Experiments show the incredible richness and robustness of an extremely simple architecture when it extracts hidden states of the HMM that feeds it with ambiguous and noisy inputs.

11:55
Self-Organizing Mappings on the Flag Manifold

ABSTRACT. A flag is a nested sequence of vector spaces. The type of the flag roughly corresponds to the sequence of dimensions of the vector spaces making up the flag. A flag manifold is a geometric object whose points parameterize all flags of a particular type in a fixed vector space. This paper provides the mathematical framework necessary for implementing self-organizing mappings on flag manifolds. Flags arise implicitly in many data analysis techniques including important examples such as wavelet, Fourier, and singular value decompositions. The proposed geometric framework in this paper enables the computation of distances between flags, the computation of geodesics between flags, and the ability to move one flag a prescribed distance in the direction of another flag. Using these operations as building blocks, we implement the SOM algorithm on a flag manifold. The basic algorithm is applied to the problem of parameterizing a set of flags of a fixed type.

12:20
Cellular Self-Organising Maps - CSOM

ABSTRACT. This paper presents CSOM, a Cellular Self-Organising Map which performs weight update in a cellular manner. Instead of updating weights towards new input vectors, it uses a signal propagation originated from the best matching unit to every other neuron in the network. Interactions between neurons are thus local and distributed. In this paper we present performance results showing than CSOM can obtain faster and better quantisation than classical SOM when used on high-dimensional vectors. We also present an application on video compression based on vector quantisation, in which CSOM outperforms SOM.

12:45
A Probabilistic Method for Pruning CADJ Graphs with Applications to SOM Clustering

ABSTRACT. We introduce a Bayesian Dirichlet-Multinomial model of the edge weights of the CADJ graph with the goal of intelligent graph pruning. As a topology representing graph, CADJ is an effective tool for cluster extraction from the learned prototypes of SOMs, but for complex data the graph must typically be pruned to elicit meaningful cluster structure. This work is a first attempt to guide this pruning in a formal modeling framework. Our model, dubbed DMPrune, earmarks edges for removal via comparisons to a novel null model and provides an internal assessment of information loss resulting from iterative removal of edges. We show that DMPruned CADJ graphs lead to clusterings comparable to the best previously achieved on highly structured real data.

14:45-16:00 Session 3

SOM: Practical Applications, Part I

14:45
Novelty detection with self-organizing maps for autonomous extraction of salient tracking features

ABSTRACT. In the image processing field, many tracking algorithms rely on prior knowledge like color, shape or even need a database of the objects to be tracked. This may be a problem for some real world applications that cannot fill those prerequisite. Based on image compression techniques, we propose to use Self-Organizing Maps to robustly detect novelty in the input video stream and to produce a saliency map which will outline unusual objects in the visual environment. This saliency map is then processed by a Dynamic Neural Field to extract a robust and continuous tracking of the position of the object. Our approach is solely based on unsupervised neural networks and does not need any prior knowledge, therefore it has a high adaptability to different inputs and a strong robustness to noisy environments.

15:10
Robust Adaptive SOMs Challenges in a Varied Datasets Analytics

ABSTRACT. The advancement of available technology in use cause the production of huge amounts of data which need to be categorised within an acceptable time for end users and decision makers to be able to make use of the data contents. Present unsupervised algorithms are not capable to process huge amounts of generated data in a short time. This increases the challenges posed by storing, analyzing, recognizing patterns, reducing the dimensionality and processing Data. Self-Organizing Map (SOM) is a specialized clustering technique that has been used in a wide range of applications to solve different problems. Unfortunately, it suffers from slow convergence and high steady-state error. The work presented in this paper is based on the recently proposed modified SOM technique introducing a Robust Adaptive learning approach to the SOM (RA-SOM). RA-SOM helps to overcome many of the current drawbacks of the conventional SOM and is able to efficiently outperform the SOM in obtaining the winner neuron in a lower learning process time. To verify the improved performance of the RA-SOM, it was compared against the performance of other versions of the SOM algorithm, namely GF-SOM, PLSOM, and PLSOM2. The test results proved that the RA-SOM algorithm outperformed the conventional SOM and the other algorithms in terms of the convergence rate, Quantization Error (QE), Topology Error (TE) preserving map using datasets of different sizes. The results also showed that RA-SOM maintained an efficient performance on all the different types of datasets used, while the other algorithms a more inconsistent performance, which means that their performance could be data type-related.

15:35
Detection of abnormal flights using fickel instances in SOM maps

ABSTRACT. In process

16:00-17:00

Invited Talk (Paulo Lisboa, U.K.)

18:30-20:00

Welcome event @ PRBB by the beach