Download pdf Advances In Self Organizing Maps Learning Vector Quantization Clustering And Data Visualization Proceedings Of The 13Th International Workshop Wsom 2019 Barcelona Spain June 26 28 2019 Alfred ebook full chapter

Advances in Self-Organizing Maps,
Learning Vector Quantization,

Clustering and Data Visualization:
Proceedings of the 13th International
Workshop, WSOM+ 2019, Barcelona,
Spain, June 26-28, 2019 Alfredo Vellido
Visit to download the full and correct content document:
https://textbookfull.com/product/advances-in-self-organizing-maps-learning-vector-qu
antization-clustering-and-data-visualization-proceedings-of-the-13th-international-wor
kshop-wsom-2019-barcelona-spain-june-26-28-2019-alfred/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...
Advances in Self-Organizing Maps, Learning Vector

Quantization, Clustering and Data Visualization:
Proceedings of the 13th International Workshop, WSOM+
2019, Barcelona, Spain, June 26-28, 2019 Alfredo
Vellido
https://textbookfull.com/product/advances-in-self-organizing-
maps-learning-vector-quantization-clustering-and-data-
visualization-proceedings-of-the-13th-international-workshop-
wsom-2019-barcelona-spain-june-26-28-2019-alfred/
Advances in Self Organizing Maps and Learning Vector

Quantization Proceedings of the 10th International
Workshop WSOM 2014 Mittweida Germany July 2 4 2014 1st
Edition Thomas Villmann
https://textbookfull.com/product/advances-in-self-organizing-
maps-and-learning-vector-quantization-proceedings-of-the-10th-
international-workshop-wsom-2014-mittweida-germany-
july-2-4-2014-1st-edition-thomas-villmann/
Business Information Systems: 22nd International

Conference, BIS 2019, Seville, Spain, June 26–28, 2019,
Proceedings, Part II Witold Abramowicz
https://textbookfull.com/product/business-information-
systems-22nd-international-conference-bis-2019-seville-spain-
june-26-28-2019-proceedings-part-ii-witold-abramowicz/
Business Information Systems: 22nd International

Conference, BIS 2019, Seville, Spain, June 26–28, 2019,
Proceedings, Part I Witold Abramowicz
https://textbookfull.com/product/business-information-
systems-22nd-international-conference-bis-2019-seville-spain-
june-26-28-2019-proceedings-part-i-witold-abramowicz/
Business Information Systems Workshops: BIS 2019
International Workshops, Seville, Spain, June 26–28,
2019, Revised Papers Witold Abramowicz
https://textbookfull.com/product/business-information-systems-
workshops-bis-2019-international-workshops-seville-spain-
june-26-28-2019-revised-papers-witold-abramowicz/
Graph Drawing and Network Visualization 26th

International Symposium GD 2018 Barcelona Spain
September 26 28 2018 Proceedings Therese Biedl
https://textbookfull.com/product/graph-drawing-and-network-
visualization-26th-international-symposium-gd-2018-barcelona-
spain-september-26-28-2018-proceedings-therese-biedl/
Swarm Intelligence 12th International Conference ANTS

2020 Barcelona Spain October 26 28 2020 Proceedings
Marco Dorigo
https://textbookfull.com/product/swarm-intelligence-12th-
international-conference-ants-2020-barcelona-spain-
october-26-28-2020-proceedings-marco-dorigo/
Big Data 7th CCF Conference BigData 2019 Wuhan China

September 26 28 2019 Proceedings Hai Jin
https://textbookfull.com/product/big-data-7th-ccf-conference-
bigdata-2019-wuhan-china-september-26-28-2019-proceedings-hai-
jin/
Evolving OpenMP for Evolving Architectures 14th

International Workshop on OpenMP IWOMP 2018 Barcelona
Spain September 26 28 2018 Proceedings Bronis R. De
Supinski
https://textbookfull.com/product/evolving-openmp-for-evolving-
architectures-14th-international-workshop-on-openmp-
iwomp-2018-barcelona-spain-september-26-28-2018-proceedings-
Advances in Intelligent Systems and Computing 976
Alfredo Vellido
Karina Gibert
Cecilio Angulo
José David Martín Guerrero Editors
Advances in
Self-Organizing
Maps, Learning Vector
Quantization, Clustering
and Data Visualization
Workshop, WSOM+ 2019, Barcelona,
Spain, June 26–28, 2019
Advances in Intelligent Systems and Computing
Volume 976
Series Editor
Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences,
Warsaw, Poland
Advisory Editors
Nikhil R. Pal, Indian Statistical Institute, Kolkata, India
Rafael Bello Perez, Faculty of Mathematics, Physics and Computing,
Universidad Central de Las Villas, Santa Clara, Cuba
Emilio S. Corchado, University of Salamanca, Salamanca, Spain
Hani Hagras, Electronic Engineering, University of Essex, Colchester, UK
László T. Kóczy, Department of Automation, Széchenyi István University,
Gyor, Hungary
Vladik Kreinovich, Department of Computer Science, University of Texas
at El Paso, El Paso, TX, USA
Chin-Teng Lin, Department of Electrical Engineering, National Chiao
Tung University, Hsinchu, Taiwan
Jie Lu, Faculty of Engineering and Information Technology,
University of Technology Sydney, Sydney, NSW, Australia
Patricia Melin, Graduate Program of Computer Science, Tijuana Institute
of Technology, Tijuana, Mexico
Nadia Nedjah, Department of Electronics Engineering, University of Rio de Janeiro,
Rio de Janeiro, Brazil
Ngoc Thanh Nguyen, Faculty of Computer Science and Management,
Wrocław University of Technology, Wrocław, Poland
Jun Wang, Department of Mechanical and Automation Engineering,
The Chinese University of Hong Kong, Shatin, Hong Kong
The series “Advances in Intelligent Systems and Computing” contains publications
on theory, applications, and design methods of Intelligent Systems and Intelligent
Computing. Virtually all disciplines such as engineering, natural sciences, computer
and information science, ICT, economics, business, e-commerce, environment,
healthcare, life science are covered. The list of topics spans all the areas of modern
intelligent systems and computing such as: computational intelligence, soft comput-
ing including neural networks, fuzzy systems, evolutionary computing and the fusion
of these paradigms, social intelligence, ambient intelligence, computational neuro-
science, artificial life, virtual worlds and society, cognitive science and systems,
Perception and Vision, DNA and immune based systems, self-organizing and
adaptive systems, e-Learning and teaching, human-centered and human-centric
computing, recommender systems, intelligent control, robotics and mechatronics
including human-machine teaming, knowledge-based paradigms, learning para-
digms, machine ethics, intelligent data analysis, knowledge management, intelligent
agents, intelligent decision making and support, intelligent network security, trust
management, interactive entertainment, Web intelligence and multimedia.
The publications within “Advances in Intelligent Systems and Computing” are
primarily proceedings of important conferences, symposia and congresses. They
cover significant recent developments in the field, both of a foundational and
applicable character. An important characteristic feature of the series is the short
publication time and world-wide distribution. This permits a rapid and broad
dissemination of research results.
** Indexing: The books of this series are submitted to ISI Proceedings,

EI-Compendex, DBLP, SCOPUS, Google Scholar and Springerlink **
More information about this series at http://www.springer.com/series/11156

Alfredo Vellido Karina Gibert
• •
Cecilio Angulo José David Martín Guerrero

•
Editors
Advances in Self-Organizing
Maps, Learning Vector
Quantization, Clustering
and Data Visualization
Workshop, WSOM+ 2019, Barcelona, Spain,
June 26–28, 2019
123
Editors
Alfredo Vellido Karina Gibert
Department of Computer Science Knowledge Engineering and Machine
UPC BarcelonaTech Learning Group (KEMLG) at Intelligent
Barcelona, Spain Data Science and Artificial Intelligence
Research Center
Cecilio Angulo UPC BarcelonaTech
Department of Automatic Control Barcelona, Spain
UPC BarcelonaTech
Barcelona, Spain José David Martín Guerrero
Departament d’Enginyeria Electrònica
Universitat de València
Burjassot, Valencia, Spain
ISSN 2194-5357 ISSN 2194-5365 (electronic)

Advances in Intelligent Systems and Computing
ISBN 978-3-030-19641-7 ISBN 978-3-030-19642-4 (eBook)
https://doi.org/10.1007/978-3-030-19642-4
© Springer Nature Switzerland AG 2020
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
The Association for Computing Machinery (ACM) has named Yoshua Bengio,
Geoffrey Hinton, and Yann LeCun as recipients of the 2018 Turing Award for their
major contribution to the development of deep neural networks as a critical com-
ponent of computing. This is a timely reminder of the renewed vitality of the
machine learning field, in which self-organizing systems have played a major role
from the 1980s not only from the perspective of data analysis, but also as in silico
models in computational neuroscience.
This book contains the peer-reviewed and accepted contributions presented at
the 13th International Workshop on Self-Organizing Maps (WSOM+2019) held at
Universitat Politècnica de Catalunya (UPC BarcelonaTech), Barcelona (Spain),
during June 26–28, 2019. WSOM+2019 is the latest in a series of biennial inter-
national conferences that started with WSOM’97 in Helsinki, Finland, with Prof.
Teuvo Kohonen as General Chairman. We would like to express our gratitude to
Prof. Kohonen for serving as Honorary Chair of WSOM+2019.
The reader will find here a varied collection of studies that testify to the vitality
of the field of self-organizing systems for data analysis. Most of them relate to the
core models in the field, namely self-organizing maps (SOMs) and learning vector
quantization (LVQ), but the workshop also catered for research in the broader
spectrum of unsupervised learning, clustering, and multivariate data visualization
problems. It is also worth highlighting that the book includes a balanced mix
of theoretical studies and applied research, covering a wide array of fields that vary
from business and engineering to the life sciences. As a result, the book should be
of interest to machine learning researchers and practitioners in general and, more
specifically, to those interested in keeping up with developments in
self-organization, unsupervised learning, and data visualization.
The book collects the work of more than 90 researchers from 18 countries, and it
is the result of a collective effort. It would not have been possible without the advice
and guidance of the international WSOM Steering Committee, and the quality
of the final selection of papers is the result of the selfless reviewing work performed
by the Program Committee members and the anonymous additional reviewers,
which enhanced the sterling work of the authors themselves. We are truly indebted
v
vi Preface
to the international researchers who agreed to participate as plenary speakers in

WSOM+2019: Prof. Paulo J. G. Lisboa (Liverpool John Moores University, UK),
Prof. Tobias Schreck (Graz University of Technology, TU Graz, Austria), Dr. Aïda
Valls (Universitat Rovira i Virgili, Spain), and Prof. Alessandro Sperduti
(Universita degli Studi di Padova, Italy). The Local Organizing Committee would
like to acknowledge the support of the UPC BarcelonaTech, the Intelligent Data
Science and Artificial Intelligence (IDEAI) Research Center at UPC and the RDLab
at the Department of Computer Science of the UPC. We also truly appreciate the
support of the sponsoring companies: Amalfi Analytics, LumenAI, and
Predict Assistance and the invaluable help provided by our postgraduate students
from the master’s and the Ph.D. in artificial intelligence programs of the UPC in the
organization of the workshop.
June 2019
Organization
Steering Committee
Teuvo Kohonen (Honorary Chairman), Finland

Marie Cottrell, France
Pablo Estévez, Chile
Timo Honkela, Finland
Jean Charles Lamirel, France
Thomas Martinetz, Germany
Erzsébet Merényi, USA
Madalina Olteanu, France
Michel Verleysen, Belgium
Thomas Villmann, Germany
Takeshi Yamakawa, Japan
Hujun Yin, UK
Program Committee
Cecilio Angulo Universitat Politècnica de Catalunya, Spain

Guilherme Barreto Universidade Federal do Ceará, Brasil
Abdel Belaid Université de Lorraine, France
Michael Biehl University of Groningen, Netherlands
Hubert Cecotti California State University, Fresno, USA
Cyril de Bodt Université catholique de Louvain, Belgium
Pablo Estévez Universidad de Chile, Chile
Jan Faigl Czech Technical University in Prague,
Czech Republic
Jérémy Fix CentraleSupélec, France
Hervé Frezza-Buet CentraleSupélec, France
vii
viii Organization
Alexander Gepperth University of Applied Sciences, HAW Fulda,

Germany
Tina Geweniger Hochschule Mittweida, Germany
Karina Gibert Universitat Politècnica de Catalunya, Spain
Yann Guermeur Laboratoire Lorrain de Recherche en
Informatique et ses Applications, France
Barbara Hammer Universität Bielefeld, Germany
Pitoyo Peter Hartono Chukyo University, Japan
Dominique Haughton Bentley University, USA
Marika Kaden Hochschule Mittweida, Germany
Michael Kirby Colorado State University, USA
Markus Koskela CSC - IT Center for Science Ltd., Finland
Jérôme Lacaille Safran Aircraft Engines, France
Jean Charles Lamirel Laboratoire Lorrain de Recherche en
Informatique et ses Applications, France
Mustapha Lebbah Université Paris 13, France
John Aldo Lee Université catholique de Louvain, Belgium
Paulo Lisboa Liverpool John Moores University, UK
Victor Lobo Universidade Nova de Lisboa, Portugal
Iván Machón Universidad de Oviedo, Spain
Raghvendra Mall Qatar Computing Research Institute, Qatar
José David Martn Universitat de València, Spain
Friedrich Melchert Fraunhofer IFF, Germany
Erzsébet Merényi Rice University, USA
Dounia Mulders Université catholique de Louvain, Belgium
Iván Olier Liverpool John Moores University, UK
Madalina Olteanu Université Paris 1, France
Sandra Ortega-Martorell Liverpool John Moores University, UK
Jaakko Peltonen Tampere University, Finland
Gilbert Peterson US Air Force Institute of Technology, USA
Andreas Rauber Vienna University of Technology, Austria
Helge Ritter Universität Bielefeld, Germany
Fabrice Rossi Université Paris 1, France
Nicolas Rougier Inria Bordeaux Sud-Ouest, France
Sascha Saralajew Porsche AG, Germany
Frank-Michael Schleif University of Applied Sciences
Würzburg-Schweinfurt, Germany
Udo Seiffert Fraunhofer IFF, Germany
César Torres-Huitzil Cinvestav Unidad Tamaulipas, México
Marc Van Hulle KU Leuven, Belgium
Alfredo Vellido Universitat Politècnica de Catalunya, Spain
Jorge Vergara Universidad de Chile, Chile
Nathalie Villa-Vialaneix INRA, France
Thomas Villmann Hochschule Mittweida, Germany
Axel Wismueller University of Rochester Medical Center, USA
Hujun Yin The University of Manchester, UK
Contents
Self-organizing Maps: Theoretical Developments

Look and Feel What and How Recurrent Self-Organizing
Maps Learn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Jérémy Fix and Hervé Frezza-Buet
Self-Organizing Mappings on the Flag Manifold . . . . . . . . . . . . . . . . . . 13
Xiaofeng Ma, Michael Kirby, and Chris Peterson
Self-Organizing Maps with Convolutional Layers . . . . . . . . . . . . . . . . . 23
Lars Elend and Oliver Kramer
Cellular Self-Organising Maps - CSOM . . . . . . . . . . . . . . . . . . . . . . . . . 33
Bernard Girau and Andres Upegui
A Probabilistic Method for Pruning CADJ Graphs
with Applications to SOM Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
Josh Taylor and Erzsébet Merényi
Practical Applications of Self-Organizing Maps, Learning Vector

Quantization and Clustering
SOM-Based Anomaly Detection and Localization
for Space Subsystems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
Maia Rosengarten and Sowmya Ramachandran
Self-Organizing Maps in Earth Observation Data Cubes Analysis . . . . . 70
Lorena Santos, Karine Reis Ferreira, Michelle Picoli, and Gilberto Camara
Competencies in Higher Education: A Feature Analysis
with Self-Organizing Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
Alberto Nogales, Álvaro José García-Tejedor, Noemy Martín Sanz,
and Teresa de Dios Alija
ix
x Contents
Using SOM-Based Visualization to Analyze the Financial

Performance of Consumer Discretionary Firms . . . . . . . . . . . . . . . . . . . 90
Zefeng Bai, Nitin Jain, Ying Wang, and Dominique Haughton
Novelty Detection with Self-Organizing Maps for Autonomous
Extraction of Salient Tracking Features . . . . . . . . . . . . . . . . . . . . . . . . . 100
Yann Bernard, Nicolas Hueber, and Bernard Girau
Robust Adaptive SOMs Challenges in a Varied Datasets Analytics . . . . 110
Alaa Ali Hameed, Naim Ajlouni, and Bekir Karlik
Detection of Abnormal Flights Using Fickle Instances in SOM Maps . . . 120
Marie Cottrell, Cynthia Faure, Jérôme Lacaille, and Madalina Olteanu
LVQ-type Classifiers for Condition Monitoring of Induction Motors:
A Performance Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
Diego P. Sousa, Guilherme A. Barreto, Charles C. Cavalcante,
and Cláudio M. S. Medeiros
When Clustering the Multiscalar Fingerprint of the City Reveals
Its Segregation Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
Madalina Olteanu and Jean-Charles Lamirel
Using Hierarchical Clustering to Understand Behavior of 3D
Printer Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
Ashutosh Karna and Karina Gibert
A Walk Through Spectral Bands: Using Virtual Reality
to Better Visualize Hyperspectral Data . . . . . . . . . . . . . . . . . . . . . . . . . 160
Henry Kvinge, Michael Kirby, Chris Peterson, Chad Eitel, and Tod Clapp
Incremental Traversability Assessment Learning
Using Growing Neural Gas Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 166
Jan Faigl and Miloš Prágr
Learning Vector Quantization: Theoretical Developments

Investigation of Activation Functions for Generalized Learning
Vector Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
Thomas Villmann, Jensun Ravichandran, Andrea Villmann,
David Nebel, and Marika Kaden
Robustness of Generalized Learning Vector Quantization Models
Against Adversarial Attacks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
Sascha Saralajew, Lars Holdijk, Maike Rees, and Thomas Villmann
Passive Concept Drift Handling via Momentum Based Robust
Soft Learning Vector Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
Moritz Heusinger, Christoph Raab, and Frank-Michael Schleif
Contents xi
Prototype-Based Classifiers in the Presence of Concept Drift:

A Modelling Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
Michael Biehl, Fthi Abadi, Christina Göpfert, and Barbara Hammer
Theoretical Developments in Clustering, Deep Learning

and Neural Gas
Soft Subspace Topological Clustering over Evolving Data Stream . . . . . 225
Mohammed Oualid Attaoui, Mustapha Lebbah, Nabil Keskes,
Hanene Azzag, and Mohammed Ghesmoune
Solving a Tool-Based Interaction Task Using Deep Reinforcement
Learning with Visual Attention . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
Sascha Fleer and Helge Ritter
Approximate Linear Dependence as a Design Method
for Kernel Prototype-Based Classifiers . . . . . . . . . . . . . . . . . . . . . . . . . . 241
David N. Coelho and Guilherme A. Barreto
Subspace Quantization on the Grassmannian . . . . . . . . . . . . . . . . . . . . . 251
Shannon Stiverson, Michael Kirby, and Chris Peterson
Variants of Fuzzy Neural Gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
Tina Geweniger and Thomas Villmann
Autoencoders Covering Space as a Life-Long Classifier . . . . . . . . . . . . . 271
Rudolf Szadkowski, Jan Drchal, and Jan Faigl
Life Science Applications

Progressive Clustering and Characterization of Increasingly
Higher Dimensional Datasets with Living Self-organizing Maps . . . . . . 285
Camden Jansen and Ali Mortazavi
A Voting Ensemble Method to Assist the Diagnosis of Prostate
Cancer Using Multiparametric MRI . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
Patrick Riley, Ivan Olier, Marc Rea, Paulo Lisboa,
and Sandra Ortega-Martorell
Classifying and Grouping Mammography Images into Communities
Using Fisher Information Networks to Assist the Diagnosis
of Breast Cancer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
Meenal Srivastava, Ivan Olier, Patrick Riley, Paulo Lisboa,
and Sandra Ortega-Martorell
Network Community Cluster-Based Analysis for the Identification
of Potential Leukemia Drug Targets . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
Adrián Bazaga and Alfredo Vellido
xii Contents
Searching for the Origins of Life – Detecting RNA Life Signatures

Using Learning Vector Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
Thomas Villmann, Marika Kaden, Szymon Wasik, Mateusz Kudla,
Kaja Gutowska, Andrea Villmann, and Jacek Blazewicz
Simultaneous Display of Front and Back Sides of Spherical SOM
for Health Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
Niina Gen, Tokutaka Heizo, Ohkita Masaaki, and Kasezawa Nobuhiko
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
Self-organizing Maps: Theoretical
Developments
Look and Feel What and How Recurrent
Self-Organizing Maps Learn
Jérémy Fix(B) and Hervé Frezza-Buet
LORIA, CNRS, CentraleSupélec, Université Paris-Saclay, 57000 Metz, France

{jeremy.fix,herve.frezza-buet}@centralesupelec.fr
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.
Keywords: Recurrent self-organizing map · Sequence processing ·

Hidden Markov Models
1 Introduction
Self-organizing maps (SOMs) or Kohonen maps, introduced in [9], is a particular
topographically organized vector quantization algorithm. It computes a mapping
from a high dimensional space to a usually one or two dimensional regular grid
with the specificity that close positions in the regular grid are associated with
close positions in the original high dimensional space. We have a pretty good
understanding of what a SOM is doing. Even if there is no energy function asso-
ciated with the Kohonen learning rule which could formally state what Kohonen
maps do actually capture (some authors actually suggested some alternative
formulations derived from an energy function, see for example [6]), we can still
pretty much see Kohonen maps as a K-means with a topology i.e. capturing the
distribution of input samples in a topographically organized fashion. As soon
as we experiment with, for example, 2D-Kohonen maps with two dimensional
input samples, we quickly face the nice unfolding of the map sometimes trapped
in some kind of local minima where there remains some twist in the map. While
our understanding of Kohonen SOMs is pretty clear, the things become more
complicated when we turn to recurrent SOMs.
Recurrent SOMs are a natural extension of SOMs when dealing with serial
inputs in order to “find structure in time” (J. Elman). This extension follows
the same principle introduced for supervised multi-layer perceptrons by [3,7] of
feeding back a context computed from the previous time step. These recurrent
SOMs are built by extending the prototype vector with an extra component
which encapsulates some information about the past. There are indeed various
proposals about the information memorized from the past, e.g. keeping only the
c Springer Nature Switzerland AG 2020
A. Vellido et al. (Eds.): WSOM 2019, AISC 976, pp. 3–12, 2020.
https://doi.org/10.1007/978-3-030-19642-4_1
4 J. Fix and H. Frezza-Buet
location of the previous best matching unit [4] or the matching over the whole
map [11]. An overview of recurrent SOMs is provided in [5]. Cellular and biolog-
ically inspired architectures have been proposed as well [8]. When the question
of understanding how recurrent SOMs work comes to the front, there are some
theoretical results that bring some answers. However, as any theoretical study,
they are necessarily limited in the questions they can address. For example, [10]
studied the behavior of recurrent SOMs by analyzing its dynamics in the absence
of inputs. As for usual SOMs for which mathematical investigations do not cover
the whole field yet [2], these theoretical results bring only a partial answer and
there is still room for experimental investigation. Despite numerous works, how
the recurrent SOMs deal with serial inputs and what they actually learn is not
obvious: “The internal model representation of structures is unclear” [5]. We
indeed lack the clear representations that we possess for understanding SOMs.
In order to tackle this issue, we focus in this paper on the simplest recur-
rent SOM where the temporal context is only the position of the best matching
unit (BMU) within the map at the previous iteration (which bears resemblance
to the SOM-SD of [4]). This simplicity comes with the ability to design spe-
cific visualizations to investigate the behavior of the map. As we shall see in
the experiments, despite this simplicity, there is still an interesting richness of
dynamics. In particular, we will investigate and visualize the behavior of this
simple recurrent SOM when inputs are provided sequentially by different hid-
den Markov models. These will illustrate the behavior of the recurrent SOM
in the presence of ambiguous observations, long-term dependencies, changing
dynamics, noise in the observations and noise in the transitions.
2 Methods
2.1 Algorithm
Let us consider a stream of inputs ξ t ∈ X available at each successive time step

t. Here, let us use X = [0, 1]. Let us consider a topological set M where unit
(sometimes called neuron) positions lie. We use a 1D map in this paper, thus
M = [0, 1] is considered with a topology induced by the Euclidean distance. The
map is made of N units, each unit is denoted by an index i ∈ I ⊂ M. Indexes
are equally spread over M, i.e. I = {0, 1/(N − 1), 2/(N − 1) · · · , 1}.
For the sake of introducing notations for our recurrent algorithm, let us start
by rephrasing the self-organizing map algorithm (SOM). Each unit i is equipped
with an input weight wi ∈ X also referred to as a prototype. When ξ t is presented
to the map, all the units compute a matching value μi = μ (wi − ξ t ), where
μ (d) = max (1 − |d|/ρμ , 0) is a linear decreasing function that reaches 0 for the
distance value ρμ 1 . The best matching unit (BMU) i can be computed here
from μi as i = argmaxi μi . It could have been computed directly as the unit
for which |wi − ξ t | is minimal, as usual SOM formulation do, but the current
formulation involving μi allows for the forthcoming extension to recurrent SOM.
1
A more classical Gaussian function could have be used as well.
Look and Feel What and How Recurrent Self-Organizing Maps Learn 5
Once i is determined consecutively to the submission of ξ t , the prototypes in

the neighborhood of i have to be updated. The strength of that update for any
wi is αh (i − i ), with α ∈ [0, 1] and h (d) = max (1 − |d|/ρh , 0). Let us stress
here that the width ρh of the learning kernel is kept constant, as opposed to usual
SOM and recursive SOM implementations where it continuously decreases.
The recurrence is added to our formulation of SOM by using context weights
ci ∈ M. They are trained in the same way as wi , exceptthat they are fed with
it−1
instead of ξ t . A context matching distribution μi = μ ci − it−1
is computed
as we did for μi . The BMU needs to be determined from both matchings. To
μ +μ
do so, each unit computes a global matching μi = i 2 i , such as the BMU
is determined as i = argmaxi μ i . The overline of μ indicates a low pass
spatial filtering with a gaussian kernel of standard deviation σk . Moreover, the
selection of the BMU is done by randomly sampling in the set of possible BMUs.
These two elements improve the algorithm in our settings where observations are
drawn from a discrete set. The whole process studied in this paper can then be
formalized into Algorithm 1. For all the experiments, we used N = 500 units, a
neighbour kernel width ρh = 0.05, a learning rate α = 0.1, a matching sensitivity
ρμ = 0.4 and a Gaussian convolution kernel standard deviation σk = 0.0125.
Algorithm 1. Architecture update at time t.

1: Get ξ t , compute ∀i ∈ I, μi = μ wi − ξ t , μi = μ ci − it−1

2: ∀i ∈ I, μi = (μi + μi )/2
3: it ∈ argmax // μ = μ ∗ k, i is taken randomly in argmax.
t i μ it−1
wi = wi +αh (i − i ) .wit−1 − ξ t
4: ∀i ∈ I t t−1 t−1 t−1
ci = ci +αh (i − i ) . ci − i
In our experiments, the inputs in X that are provided at each time step
are generated from a Hidden Markov Model (HMM). The HMM has a finite
set S = {s0 , s1 , · · ·} of states. Each state is an integer (i.e. S ⊂ N). At each
time step, a state transition is performed according to a transition matrix. In
the current state st , the observation is sampled from the conditional probability
P (ξ | st ), defined by the observation matrix of the HMM. Different states of the
HMM may provide a similar observation. In this case, the recursive architecture
is expected to make the difference between such states in spite of the observation
ambiguity. In other words, the current BMU it value is expected to represent the
actual st even if several other states could have provided the current input ξ t .
2.2 Representations
Algorithm 1 can be executed with any dimension for M without loss of generality.
Nevertheless, we use 1D maps (M = [0, 1]) for the sake of visualization. Weights
wi are in X = [0, 1] as well. They can be represented as a gray scaled value, from
black (0) to white (1). In the bottom left part of Fig. 1, the background of the
chart is made of wit , with t in abscissa and i in ordinate. On this chart, red curves
are also plotted. This is done when the HMM is deterministic (and thus cycling
through its states, visiting s0 , s1 , · · · , sp−1 , s0 , s1 , · · · ). If the state sequence that
is repeated throughout the experiment has a length p (p = 10 in experiment
of Fig. 1), p red curves are plotted on the chart. For 0 ≤ k < p, the kth red
curve links the points {(t, it ) | t mod p = k}. The curves show the evolution of
the BMU position corresponding to each of the p states throughout learning.
From left to right in that chart in Fig. 1, some red curves are initially overlaid
before getting progressively distant. Such red curves splits show a bifurcation
since the map allocates a new place on M for representing a newly detected
HMM state. This allocation has a topography since the evolution is a split and
then a progressive spatial differentiation of the state positions.
Let us take another benefit from using 1D maps and introduce an original
representation of both w and c weights. This representation is referred to as a
protograph in this paper. It consists of a circular chart (see three of them on top
left in Fig. 1). The gray almost-closed circle represents M = [0, 1]. At time step
t, one can plot on the circle the two weights related to it . First weight, related to
the input, is w (it ), which is a value in X to which a gray level is associated. This
is plotted as a gray dot with the corresponding gray value, placed on the circle
at position it . The second weight to be represented for it is c (it ), related to the
recurrent context, which is a position in M and thus a position on the circle.
c (it ) is represented with an arrow, starting from position c (it ) on the circle and
pointing at it on the circle, where the dot representing w (it ) is actually located.
This makes a dot-arrow pair for it . The full protograph at time t plots the dot-
arrow pairs (w (i ), c (i )) for the 50 last steps. The third protograph in Fig. 1
seems to contain only 10 dot-arrow pairs since many of the 50 ones are identical
to others. This last protograph corresponds to an organized map, it reveals the
number of states visited by the HMM (number of dots), where they are encoded
in the map (dot positions), which observation each state provides (dot colors),
and the state sequence driven by the HMM transitions (follow the arrows from
one state to another). Making movies from the succession of such protographs
unveils the dynamics of the organization of spatio-temporal representations in
the map. The splits and separation mentioned for the red curves is then visible
as a split of one dot into two dots that slide afterwards away one from the other.
Movies of the experiments are available online2 .
2.3 Evaluation
The representations presented so far enables to unveil the inner dynamics of

a single run. Nevertheless, the ability of the architecture to encode the hidden
states of the HMM providing the inputs needs to be measured quantitatively
from several runs (athousand in our experiments).
t−1 t−1 At time step t, let us store
the dataset Dt = it−99
, st−99
, · · · , i , s , (it
, st
) that is a 100-sized
sliding window containing the last observed BMU position/HMM state pairs.
2
http://www.metz.supelec.fr/∼fix jer/recsom1D.
If the map encodes the HMM states with a dedicated BMU position, each
observed BMU position must be paired with a single state. In this case, Dt
can be viewed as a set of samples of a function from M to S. To check this
property for the map at time t, a supervised learning process is performed from
Dt , that is viewed here as an input/output pairs container. As S ⊂ N, this is
a multi-class learning problem. A basic bi-class decision stump is used in this
paper (i.e. a threshold on map position values makes the decision), adapted to
the multi-class problem thanks to a one-versus-one scheme. Let us denote by χt
the classification error rate obtained on Dt (i.e. the empirical risk). The value
χt is null when one can recover the state of the HMM from the position of the
BMUs collected during the 100 steps. It is higher when a small contiguous region
of the map is associated with several HMM states.
In our experiment, χt is computed every 100 steps in a run. As previously
said, 1000 runs are performed in order to compute statistics about the evolution
of χt as the map gets organized. At each time step t, only the best 90% of the
1000 χt are kept. The evolution curve, as reported in the right of Fig. 1, plots
the upper and lower bounds of these 900 values, as well as their average. There
are indeed less than 10% of the runs for which the map does not properly self-
organize. A deeper investigation of this phenomenon is required, but it is out
of the scope of the present paper, which is focused on the dynamic of the self-
organization when it occurs. This is why the corresponding runs are removed
from the performance computation.
3 Results
As mentioned in Sect. 2.1, the serial inputs ξ t are observations provided by the
successive states of a HMM. Let us use a comprehensive notation for the HMMs
used in our experiments. Observations are in X = [0, 1] as previously stated and
6 specific input values are represented by a letter (A = 0, B = 0.2, C = 0.4, D =
0.6, E = 0.8, F = 1). The HMM denoted by AABFCFE is then a 7-state HMM
for which s0 provides observation A, s1 provides A as well, s2 provides B, ... s6
provides E. The states are visited from s0 to s6 periodically. In this particular
HMM, (s0 , s1 ), as well as (s3 , s5 ) are ambiguous since they provide the same
observation (A and F) as an input to the recurrent SOM. When a state provides
an observation uniformly sampled in [0, 1], it is denoted by ∗. The notation
σ
ABCD EF means that values for both s2 and s3 are altered by an additive normal
noise with standard deviation σ. Last, the notation ABCpq DEF means that the
HMM is made of two periodical HMMs ABC and DEF, with random transitions
from any of the state of ABC to any of the state of DEF with a probability p.
Random transitions from DEF to ABC occurs similarly with a probability q.
3.1 Ambiguous Observations
In order to test the ability of the recurrent SOM to deal with ambiguous obser-
vations, we consider the HMM ABCEFEDCB, i.e. a HMM with 10 states and 6
observations. There are 8 states which provide an observation that is ambiguous

(the state cannot be identified given only the current observation). The recurrent
SOM receives observations during 5000 time steps, i.e. 500 presentations of the
full sequence. A single run is depicted on the left of Fig. 1. The SOM initially
captures the individual observations, irrespective of their context (at t = 200,
approximately 6 units are the BMUs), and then, ambiguous observation begin
electing their own BMUs and the recurrent prototypes begin to reflect the con-
text of the observations. This splitting of the winning positions ultimately leads
to one distinct BMU for each state of the HMM and, at t = 5000, the structure
of the HMM can be uncovered from the weights (w (i ) , c (i )) as displayed in
the respective protograph. Red curves show how the 6 areas in the map insen-
sitive to the context split to build the expected appropriate 10 areas. Running
the same experiment for 1000 runs indicates that the ability of the algorithm
to identify the structure of the HMM is statistically significant (right of Fig. 1).
The observations produced by this HMM could be easily disambiguated, taking
into account the observation of the previous time step. In the next experiment,
we study longer term dependencies.
0.5
0.2 0.2 0.2 decision stump risk
0.4 0.4 0.4

0.4
0.0 0.0 0.0
1.0 1.0 1.0

0.3
0.6 0.6 0.6
0.8 0.8 0.8

1.0
0.2
0.8
0.6
0.1
0.4
0.2
0.0
0.0
1000 2000 3000 4000 5000 0 2000 4000 6000 8000 10000
step
Fig. 1. Observations from ABCDEFEDCB. The protographs are recorded at t =

200, 500, 5000.
3.2 Long Term Dependencies

In the second experiment, the algorithm receives observations from the HMM
AAAAAAAF. This experiment seeks to test if the algorithm can capture long-term
dependencies. Indeed, this HMM produces exactly the same observation A for
a fixed number of steps before outputting F which seeds the ambiguity of A.
The experiment is run for 5000 steps, i.e. 625 repetitions of the full sequence.
The results of a single run are displayed on the left of Fig. 2. Initially, the SOM
captures the two observations A and F independently of the context (see the plot
of the prototypes on the top left of the figure). Then, we observe several units
specializing to the observation of A in a context dependent manner. The logic
of the propagation of the context can be appreciated from the split of the red
curves. The first state producing a A to be clearly identified is the one associated
with the unit with the smallest position (the node shown in back on the first
protograph around position 0.3). This is the simplest to be identified because it

is the state just after the one outputting F. Then, the dependence on the context
propagates through all the previous steps. The c (i ) weights are continuous over
M, as for usual weights in SOMs. In the specific run of Fig. 2, when we consider
the black dots counterclockwise, the arrow origins c (i ) progress clockwise, i.e.
c (i ) is a monotonously decreasing function. Running this experiment on 1000
independent trials reveals that the algorithm is able to capture the structure of
this HMM (see Fig. 2, right).
0.5
0.2 0.2 0.2 decision stump risk
0.4 0.4 0.4

0.4
0.0 0.0 0.0
1.0 1.0 1.0

0.3
0.6 0.6 0.6
0.8 0.8 0.8

1.0
0.2
0.8
0.6
0.1
0.4
0.2
0.0
0.0
1000 2000 3000 4000 5000 0 2000 4000 6000 8000 10000
step
Fig. 2. Observations from AAAAAAAF. The protographs are recorded at t =

40, 1000, 5000.
3.3 Adapting to a Changing Dynamics
In this third experiment, the algorithm receives observations from the HMM
ABCDEFEDCB for the 10000 first steps and then from the HMM ABCBAFEDEF for
the last 10000 steps. The prototypes obtained at t = 10000 and t = 20000 as well
as the evolution of the observation weights and winner locations are displayed
on the left of Fig. 3 for a single run. The algorithm successfully recovers the
structure of the two HMM. Analyzing the red curves at the time the second
HMM is presented is illuminating. One can note there is a reuse of the previously
learned prototypes and some adaptation of the prototypes. Indeed, there was a
single BMU responsive for a F (white node on the first protograph) for the first
sequence which splits and two BMUs are now responsive for a F for the second
sequence, which makes sense given the second HMM has two different states
producing the observation F. The same comment holds for the BMUs when the
observation A is produced by the HMM. On the contrary, while two BMUs had
observation prototypes w close to a C and D during the first training period, only
one BMU is remaining for each C and D after learning with the second HMM.
The performance of the algorithm ran for 1000 independent trials is shown on
the right of Fig. 3. Similarly to the first experiment, it takes around 5000 steps
to learn the sequence. At the time the HMM is changed, there is a degradation
in the performances that quickly drops.
0.2 0.2 0.5

decision stump risk
0.4 0.4
0.4
0.0 0.0
1.0 1.0 0.3
0.6 0.6
0.2
0.8 0.8
1.0
0.8 0.1
0.6
0.4
0.2 0.0
0 2500 5000 7500 10000 12500 15000 17500 20000
0.0 step
0 2500 5000 7500 10000 12500 15000 17500 20000
Fig. 3. Observations from ABCDEFEDCB for the first 10000 steps and then from
ABCBAFEDEF. The protographs are recorded at t = 10000, 20000.
3.4 Noisy Observations
We now perform an experiment to test the robustness of the algorithm in

the presence of noise in the observations. This experiment involves the HMM
0.05
BCDEDC , i.e. each observation is perturbed with a normal noise of standard
deviation 0.05. Given that the unperturbed observations are separated by 0.2,
a normal noise of standard deviation 0.05 leads to slightly overlapping obser-
vations. On the left of Fig. 4, one can recognize the 6 states identified by the
algorithm. The blur in the representation of the prototypes comes from the fact
that the circles arranged along the ring are displayed with a transparency pro-
portional to the winning frequency of the displayed BMU. Given the observations
are noisy, there is jitter in the location of the BMUs but still, the elected BMUs
for each state of the HMM remains in a compact set. The sequence B, C, D, E, D, C
tends to elicit BMUs in positions around 0.5, 0.65, 0.1, 0.35, 0.2 and 0.8. Run-
ning the experiment on 1000 runs, the performance χt decreases almost down
to 0.0 as shown on the right of Fig. 4. This confirms that the locations that are
BMUs for a given HMM state are indeed compact sets. Finally, while the absence
of noise and the linear neighborhood function kept the BMUs confined within
the map in the previous experiments, the presence of noise in this experiment
ultimately leads to populate all the map; all the positions within the map tend
to be recruited to encode the HMM.
3.5 Perturbed by a Noise State

In the last experiment, we consider a challenging HMM ABCDEFEDCBpq ∗ , with
p = 0.03 and q = 0.1. This HMM is based on the sequence ABCDEFEDCB. There
is a probability p for the HMM to jump from any of the states ABCDEFEDCB to
the state we denote ∗ which emits a uniformly distributed observation. There
is also a probability q to jump from the state ∗ back to one of the states in
0.2 0.5
decision stump risk
0.4
0.4
0.0
1.0 0.3
0.6
0.2
0.8
1.0
0.8 0.1
0.6
0.4
0.2 0.0
0 2500 5000 7500 10000 12500 15000 17500 20000
0.0 step
0 2500 5000 7500 10000 12500 15000 17500 20000
0.05
Fig. 4. Observations from BCDEDC . The protograph is recorded at t = 20000.
ABCDEFEDCB. The clean sequence of observations from ABCDEFEDCB is therefore

regularly corrupted with a uniform noise. This challenging HMM can mimic for
example a temporal disruption of sensors. With p = 0.03, there is a probability
(1 − p)9 = 0.76 to completely unroll the sequence ABCDEFEDCB when starting
from A. When the HMM is in the state ∗, it stays in this state for 1q = 10 steps
in average. The results of a single run are displayed on the left of Fig. 5. The
structure of the HMM, without the noisy state, can be recognized from the plot
of the prototypes on the top of the figure (if we omit the BMUs just before i = 0.2
and just after i = 0.8). It should be noted that the structure of the algorithm does
not allow it to capture the noisy state and the latter is therefore filtered by the
algorithm. Running the experiment for 1000 runs indicates that the ability of the
algorithm to capture the structure of the clean HMM is statistically significant,
as shown on the right of Fig. 5. For computing the statistics on Fig. 5, the samples
labelled with ∗ are removed from the dataset Dt . However, this still does not
0.2 0.5
decision stump risk
0.4
0.4
0.0
1.0 0.3
0.6
0.2
0.8
1.0
0.8
0.1
0.6
0.4
0.2 0.0
0 10000 20000 30000 40000 50000
0.0 step
0 10000 20000 30000 40000 50000

Fig. 5. Observations from ABCDEFEDCBpq ∗ . The protograph is recorded at t = 20000.
lead to a perfect classification. Indeed, when the HMM is back from the noisy
state ∗, it sometimes requires two successive observations to identify in which
state the HMM is. This explains why χt is not null.
4 Conclusion
This paper presents an empirical approach of recurrent self-organizing maps by
introducing original representations and performance measurements. The exper-
iments show how spatio-temporal structure gets organized internally to retrieve
the hidden states of the external process that provides the observations. An area
of the map associated with an observation splits into close areas when obser-
vation ambiguity is detected, and then areas get progressively separated onto
the map. Unveiling the emergence of such a complex and continuous behavior,
from both the SOM-like nature of the process and a simple re-entrance, is the
main result of this paper. Such a simple architecture also shows robustness to
temporal and spatial damages in the input series, as well as the ability to deal
with deep time dependencies while the recurrence only propagates previous step
context. Forthcoming work will consist in using such recurrent maps in more
integrated multi-map architecture, as started in [1].
Acknowledgement. This work is supported by the European Interreg Grande

Région/Région Grand-Est project GRONE.
References
1. Baheux D, Fix J, Frezza-Buet H (2014) Towards an effective multi-map self orga-
nizing recurrent neural network. In: ESANN, pp 201–206
2. Cottrell M, Fort J, Pags G (1998) Theoretical aspects of the SOM algorithm.
Neurocomputing 21(1):119–138
3. Elman JL (1990) Finding structure in time. Cognit Sci 14:179–211
4. Hagenbuchner M, Sperduti R, Tsoi AC, Member S (2003) A self-organizing map
for adaptive processing of structured data. IEEE Trans Neural Netw 14:491–505
5. Hammer B, Micheli A, Sperduti A, Strickert M (2004) Recursive self-organizing
network models. Neural Netw 17(8–9):1061–1085
6. Heskes T (1999) Energy functions for self-organizing maps. In: Oja E, Kaski S
(eds) Kohonen maps. Elsevier Science B.V, Amsterdam, pp 303–315
7. Jordan MI (1996) Serial order: a parallel distributed processing approach. Technical
reporty 8604, Institute for Cognitive Science, University of California, San Diego
8. Khouzam B, Frezza-Buet H (2013) Distributed recurrent self-organization for
tracking the state of non-stationary partially observable dynamical systems. Biol
Inspired Cognit Archit 3:87–104
9. Kohonen T (1982) Self-organized formation of topologically correct feature maps.
Biol Cybern 43(1):59–69
10. Tiňo P, Farkaš I, van Mourik J (2006) Dynamics and topographic organization of
recursive self-organizing maps. Neural Comput 18(10):2529–2567
11. Voegtlin T (2002) Recursive self-organizing maps. Neural Netw 15(8–9):979–991
Self-Organizing Mappings
on the Flag Manifold
Xiaofeng Ma, Michael Kirby(B) , and Chris Peterson
Colorado State University, Fort Collins, CO 80523, USA

xiaofeng.ma@rams.colostate.edu,
{michael.kirby,christopher2.peterson}@colostate.edu
Abstract. A flag is a nested sequence of vector spaces. The type of the

flag is determined by the sequence of dimensions of the vector spaces
making up the flag. A flag manifold is a manifold whose points param-
eterize 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 for instance in 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 build-
ing 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.
Keywords: Self-Organizing Mappings · SOM · Flag manifolds ·

Geodesic · Visualization
1 Introduction
Self-Organizing Mappings (SOMs) were introduced as a means to see data in

high-dimensions [7–10]. This competitive learning algorithm effectively trans-
ports the notion of proximity in the data space to proximity in the index space;
this may in turn be endowed with its own geometry. This tool has now been
widely applied and extended [4]. The goal of the SOM algorithm is to produce
a topology preserving mapping in the sense that points that are neighbors in
high-dimensional space are also represented as neighbors in the low-dimensional
index space.
The geometric framework of the vanilla version of the SOM algorithm is
Euclidean space. In this setting, the distance between points is simply the stan-
dard 2-norm of the vector difference. The movement of a center towards a pat-
tern takes place on a line segment in the ambient space. The only additional
ingredient to the algorithm is a metric on the index space.
c Springer Nature Switzerland AG 2020

A. Vellido et al. (Eds.): WSOM 2019, AISC 976, pp. 13–22, 2020.
https://doi.org/10.1007/978-3-030-19642-4_2
14 X. Ma et al.
Motivated by the subspace approach to data analytics we proposed a version

of SOM using the geometric framework of the Grassmannian [2,14–16]. This sub-
space approach has proven to be effective in settings where you have a collection
of subspaces built up from a set of patterns drawn from a given family. Given
one can compute distances between points on a Grassmannian, and move one
point in the direction of another, it is possible to transport the SOM algorithm
on Euclidean space to an SOM algorithm on a Grassmannian [6,11].
An interesting structure that generalizes Grassmannians and encodes addi-
tional geometry in data is known as the flag manifold. Intuitively, a point on a
flag manifold is a set of nested subspaces. So, for example, given a data vector,
a wavelet transform produces a set of approximations that live in nested scaling
subspaces [5]. The nested sequence of scaling subspaces is a flag and corresponds
to a single point on an appropriate flag manifold. Alternatively, an ordered basis,
v1 , v2 , . . . , vk for a set of data produced by principal component analysis induces
the flag S1 ⊂ S2 ⊂ · · · ⊂ Sk where Si is the span of v1 , . . . , vi . In this paper
we extend SOM to perform a topology preserving mapping on points that cor-
respond to nested subspaces such as those arising, for instance, from ordered
bases or wavelet scaling spaces. To accomplish this we show how to compute the
distance between two points on a flag manifold, and demonstrate how to move
a flag a prescribed distance in the direction of another. Given these building
blocks, we illustrate how one may extend SOM to the geometric framework of a
flag manifold.
This paper is outlined as follows: In Sect. 2 we provide a formal definition of
the flag manifold and illustrate with concrete examples. In Sect. 3 we introduce
the numerical representation of flag manifolds. Here we indicate explicitly how
distances can be computed between flags, and further, how a flag can be moved
in the direction of another flag. In Sect. 4 we put the pieces together to realize
the SOM algorithm on flag manifolds. We demonstrate the algorithm with a
preliminary computational example. Finally, in Sect. 5 we summarize the results
of the paper and point towards future directions of research.
2 Introduction to Flag Manifold with Data Analysis

Examples
Let us first introduce the flag manifold. A f lag of type (n1 , n2 , . . . , nd ; n) is a

nested sequence of subspaces in Rn where {0} V1 V2 · · · Vd = Rn ,
j
dim Vj = Σi=1 ni and n1 + n2 + · · · + nd = n. We let F L(n1 , n2 , . . . , nd ; n) denote
the flag manifold whose points parameterize all flags of type (n1 , n2 , . . . , nd ; n).
As a special case, the flag of type (1, 1, · · · , 1; n) is referred to as a full flag
and F L(1, 1, · · · , 1; n) is the full flag manifold in Rn . Figure 1 illustrates the
nested structure of the first three low-dimensional elements comprising a full
flag in Rn . A flag of type (k, n − k; n) is simply a k–dimensional subspace of Rn
(which can be considered as a point on the Grassmann manifold Gr(k, n)). Hence
F L(k, n − k; n) = Gr(k, n). The Grassmannian-SOM algorithm is developed in
[6,11]. The idea that the flag manifold is a generalization of the Grassmann
SOM on the Flag Manifold 15
Fig. 1. Illustration of a nested sequence of subspaces corresponding to a point on the

flag manifold F L(1, 1, · · · , 1; n).
manifold will be utilized later to introduce the geodesic formula on the flag
manifold. The nested structure inherent in a flag shows up naturally in the
context of data analysis.
1. Wavelet analysis: Wavelet analysis and its associated multiresolution repre-
sentation produces a nested sequence of vector spaces that approximate data
with increasing resolution [1,12,13]. Each scaling subspace Vj is a dilation
of its adjacent neighbor Vj+1 in the sense that if f (x) ∈ Vj then a reduced
resolution copy f (x/2) ∈ Vj+1 . The scaling subspaces are nested
· · · ⊂ V2 ⊂ V1 ⊂ V0 ⊂ V−1 ⊂ · · ·
and in the finite dimensional setting can be considered as a point on a flag
manifold. The flag SOM algorithm provides a means to visualize relationships
in a collection of discrete wavelet transforms and organize the corresponding
sequences of nested subspaces in a coherent manner via a low-dimensional
grid.
2. SVD basis of a real data matrix: Let X ∈ Rn×k be a real data matrix con-
sisting of k samples in Rn . Let U ΣV T = X be the thin SVD of X. The
columns of the n-by-d orthonormal matrix U is an ordered basis for the
column span of X. This basis is ordered by the magnitude of the singular
values of X. This order provides a straightforward way to associate to U a
point on a flag manifold. If U = [u1 |u2 | . . . |ud ] then the nested subspaces
span([u1 ]) span([u1 |u2 ]) · · · span([u1 | · · · |ud ]) Rn is a flag of type
(1, 1, . . . , 1, n − d; n) in Rn . After we introduce the distance metric on the flag
manifold in Sect. 3.2, one could consider computing the distance between two
flags, perhaps derived from a thin SVD of two different data sets, which takes
the order of the basis into consideration.
3 Numerical Representation and Geodesics

A point in the vector space Rn can be naturally represented by an n × 1 vector.
For a more abstract object like a Grassmann or flag manifold, we need a way
to represent points in such a way that we can do computations. In this section,
we describe how we can represent points and we describe how to determine and
express geodesic paths between points. Note that in this paper we are using exp
and log to denote the matrix exponential and the matrix logarithm.
16 X. Ma et al.
3.1 Flag Manifold

The flag manifold F L(n1 , n2 , . . . , nd ; n) consists of the set of all flags of type
(n1 , n2 , . . . , nd ; n). The presentation in [3] describes how to view the Grassmann
manifold Gr(k, n) as the quotient manifold SO(n)/S(O(k)×O(n−k)). Similarly,
we can view F L(n1 , n2 , . . . , nd ; n) as the quotient manifold SO(n)/S(O(n1 ) ×
O(n2 ) × · · · × O(nd )) where n1 + n2 + · · · + nd = n. Let P ∈ SO(n) be an n-by-n
orthogonal matrix, the equivalence class [P ], representing a point on the flag
manifold, is the set of orthogonal matrices
⎧ ⎛ ⎞ ⎫
⎪
⎪ P1 0 · · · 0 ⎪
⎪
⎪
⎨ ⎜ 0 P2 · · · 0 ⎟ ⎪
⎬
⎜ ⎟
[P ] = P ⎜ . . . ⎟ : P i ∈ O(ni ) , n1 + n2 + · · · + nd = n .
⎪
⎪ ⎝ .. . . .. ⎠ ⎪
⎪
⎪
⎩ ⎪
⎭
0 ··· Pd
It is well known that the geodesic paths on SO(n) are given by exponential flows
Q(t) = Q exp(tA) where A ∈ Rn×n is any skew symmetric matrix and Q(0) = Q.
Viewing F L(n1 , n2 , . . . , nd ; n) as a quotient manifold of SO(n), one can show
that geodesics on SO(n) continue to be geodesics on F L(n1 , n2 , . . . , nd ; n) as
long as they are perpendicular to the orbits generated by S(O(n1 ) × O(n2 ) ×
· · · × O(nd )) (for a derivation on a Grassmann manifold, see [11]). This leads
one to conclude that the geodesic paths on F L(n1 , n2 , . . . , nd ; n) are exponential
flows:
P (t) = P exp(tC̃) (1)
where C̃ is any skew symmetric matrix of the form

⎛ ⎞
0n1 ∗
⎜ 0n2 ⎟
⎜ ⎟
C̃ = ⎜ . ⎟ , 0ni = 0ni ×ni .
⎝ .. ⎠
−∗T 0nd
3.2 Geodesic and Distance Between Two Points on Flag Manifold

By Eq. (1), one may trace out the geodesic path on a flag manifold emanating
from P in the direction of C̃. In this section we utilize Eq. (1) to solve the inverse
problem:
Given two points Q1 , Q2 ∈ SO(n), whose equivalence classes [Q1 ], [Q2 ] rep-
resent flags of type (n1 , n2 , . . . , nd ; n), obtain a factorization
Q2 = Q1 · exp(H) · M (2)
for H and M where H and M are constrained to be of the form
⎛ ⎞ ⎛ ⎞
0n1 ∗ M1 0 · · · 0
⎜ 0n2 ⎟ ⎜ 0 M2 · · · 0 ⎟
⎜ ⎟ ⎜ ⎟
H=⎜ . ⎟ and M = ⎜ . . . .. ⎟
⎝ .. ⎠ ⎝ .. . . ⎠
−∗T
0nd 0 ··· Md
SOM on the Flag Manifold 17
where H is skew symmetric, Mi ∈ O(ni ), and M ∈ SO(n). The distance

between [Q1 ] and [Q2 ] along the geodesic given by H is

l λ2
d([Q1 ], [Q2 ]) = Σi=1 (3)
i
where the λi ’s are the distinct singular values of H.
Fig. 2. Illustration of Eq. (2). The vertical lines represents the equivalence classes [Q1 ]
and [Q2 ] respectively. Q1 is mapped to an element in [Q2 ] by right multiplication with
exp(H) which is then sent to Q2 by multiplying with M .
Equation (2) can be interpreted in the following way. First, we map Q1 to a

representative in [Q2 ] via the geodesic determined by the velocity matrix H. Sec-
ond, we map this element in [Q2 ] to Q2 via the matrix M . Figure 2 is a pictorial
illustration of the idea behind Eq. (2). For F L(k, n − k; n) i.e. the Grassman-
nian Gr(k, n), one can solve for H analytically. Please see [3] for details. For the
more general case, we will present an iterative algorithm to obtain a numerical
approximation of H and M in Sect. 3.3. Before we proceed to the algorithm, let
us further simplify Eq. (2) by letting Q = QT1 Q2 . This allows us to rewrite (1) as
Q = exp(H) · M (4)
Here we define W as the vector space of all n-by-n skew symmetric matrices. Let
p = (n1 , n2 , . . . , nd ; n). We define Wp to be the set of all block diagonal skew
symmetric matrices of type p and its orthogonal complement Wp⊥ in W, i.e.
⎛ ⎞ ⎛ ⎞
G1 ··· 0 0n1 ∗
⎜ .. ⎟}, W ⊥ = {H ∈ W | H = ⎜ ⎟
Wp = {G ∈ W | G = ⎝ ... ..
. . ⎠ p ⎝ ..
. ⎠}.
0 ··· Gd −∗ T
0nd
where, by definition, Gi ∈ Rni ×ni is skew symmetric for all i. Instead of solving
Eq. (4) directly, we propose to solve the following alternative equation:
Q = exp(H) · exp(G) (5)

Another random document with
no related content on Scribd:
reduced outer end of the arm, and a linch-pin is driven through the
arm beyond it.
An improvement on this kind of axle is when the collar at the upper
end or shoulder is made solid by welding, and a screw nut with a
linch-pin through it is substituted for the collar and linch-pin. These
nuts are commonly made six-sided, with a mortise or slot for the
linch-pin through each side, in order to afford greater facility for
adjustment. In all other particulars this axle is the same as the last,
except that it is occasionally case-hardened to prevent wear and
friction.
In travelling, these axles require to be fresh greased every two or
three days, and the trouble thus caused is very considerable,
besides the risk of omission, in which case the axle is likely to be
entirely spoiled.
The commonest kind of oil axle is called the “mail,” because the
peculiar mode of fastening was first used in the mail coaches. The
arm is not conical, but cylindrical, in the improved kind. At the
shoulder of this axle a solid disc collar is welded on for the box to
work against. Behind this shoulder collar revolves a circular flange-
plate of wrought iron, pierced with three holes corresponding with
holes in the wheel from front to back, through which long screw-bolts
are driven, and their nuts screwed sufficiently tight against the
circular flange-plate to allow easy motion. The wheel, when in
motion, thus works round the shoulder collar, while the flange-plate
secures it against coming off. This is not neat or accurate, but it is
simple and secure, and no nut or linch-pin is required to the axle in
front, while the front of the nave can be entirely covered in. When
screwed up for work, a washer of thick leather is placed between the
shoulder collar and the box, and another between the shoulder collar
and the circular disc, which extends over the whole surface of the
back of the nave. The box of this axle is of cast iron. The front is
closed with a plate of metal, between which and the end of the axle-
arm a space is left of about 1 inch as a reservoir for oil, which is
poured in through a tube passing through the nave of the wheel and
closed by a screw pin. At the back of the box there is a circular
reservoir for oil, ¾ inch in depth and ½ inch wide. When the wheel is
in motion the revolving of the box keeps the lubricating material in
circulation between the two reservoirs; any portion getting below the
arm at the shoulder gradually works its way out and is wasted. The
oil in the back reservoir does not waste by leakage so rapidly as that
in the front; but when the leather washer becomes saturated with
water the oil is liable, by reason of its lightness, to float on the water
in or about the washer, and thus get wasted.
This axle requires frequent examination when very much in use;
but as it is neat in appearance, and under ordinary circumstances
tolerably safe in working, and is not very expensive, it is much used.
Both axle-box and axle-arm are case-hardened.
The other kind of axle used by carriage-builders is that known as
“Collinge’s Patent.” The original intention of the inventor was to make
it a cylindrical arm, with the box running round it against a coned
shoulder, and secured by a coned nut in front; but, as it was found in
practice that a leather washer was necessary at the shoulder to
prevent jarring, this part of the plan was abandoned.
The commonest form of this axle now in use consists of a
cylindrical arm with a broad shoulder collar. The box is of cast iron,
and the back of it is similar to that of the mail axle before described.
The front of it has a rebate cut in the box to receive a small conical
collar and the screw of an oil cap. The arm of the axle is turned down
in the lathe to two-thirds of the total thickness from the point where
the rebate of the box begins. A flat side is filed on this reduced
portion, and along it is made to slide a small collar of gun metal, with
a conical face in the interior to fit against the coned interior of the
rebate in the box. Against this collar, technically called the “collet,” a
nut of gun metal is screwed, and against that again a second nut of
smaller size, with a reversed thread, is tightly fixed. These two nuts,
thus screwed in different directions, become as firm as though they
were part of the axle itself, and no action of the wheel can loosen
them, because the collet, which does not turn, removes all friction
from them. But, as a further security, the end of the axle-arm projects
beyond the farthest nut, and is drilled to receive a spring linch-pin.
Over all a hollow cap of gun metal is screwed into the end of the box.
This contains a supply of oil for lubricating purposes.
When the wheel is in motion the oil is pumped upwards from the
cap and passes along the arm to the back reservoir, constantly
revolving round the cap with the wheel. If the cap be too full of oil—
that is, if the summit of the column of oil in the cap be at a horizontal
level above the leakage point at the shoulder—it will pump away
rapidly, and be wasted till it comes to the level of the leak, where it
will be economically used. It is essential to the perfection of an oil
action that the oil should not be permanently above the level of the
leak, but that small portions should be continually washing up into
that position by the action of the wheel in turning.
In order to insure their greater durability and freedom from friction
these axles and their boxes are always case-hardened, i.e. their
rubbing surfaces are converted into steel to a trifling depth by the
process of cementation with animal charcoal for about two hours,
when they are plunged into water. The boxes are ground on to the
arms with oil and emery, either end being applied alternately, until a
true fit between the two is accomplished.
The mode in which oil acts as a lessener of friction is by its being
composed of an infinite number of movable globules, over which the
fixed surfaces of the arm and box roll without causing that friction
and wearing away which would be the result of the two iron surfaces
worked together without any lubricant. This saving in the wear and
tear of the axle-arm is accomplished by the destruction of the oil.
From this we deduce that the greater the mass of oil or grease used
the longer will the axle run, and in order to facilitate this as much as
possible there should be so much space left between the bearing
surfaces of the arm and the box as will allow of a film of oil to be
between them.
A highly polished surface is desirable in an axle and box, as the
bearing is more perfect and true. A rough surface is a surface of
sharp angles, which will pierce through the oil and cause friction by
contact.
To guard against the axle running dry, the arm is reduced in
thickness at the centre for about an inch to allow a lodgment for the
oil, and in the process of working this constitutes a circular pump,
which draws up the oil from the front cap and distributes it over the
area of the arms. But this, of course, will soon run dry, so that the
best remedy to prevent the oil being exhausted and the sticking of
the axle-arm in the box is careful attention.
A danger arising from careless fitting is the introduction of grit into
the box. This grit is composed of small grains of silex, which is very
much harder than iron or steel; the consequence is that it cuts and
scores the bearing surfaces in all directions, and keys them firmly
together, so that it is sometimes necessary to break the box to
pieces in order to get it off the arm.
A patent was taken out to remedy these defects by casting three
longitudinal triangular grooves in each box. The advantages gained
by this are, that if grit gets in it finds its way to the bottom of the
grooves and does not interfere with the action of the wheel, and,
moreover, the grooves keep up a constant surface of oil in contact
with the arm, instead of trusting to the mere capillary attraction. This
does not interfere with the bearing surface in any marked degree.
In order that the axle shall be perfect the following considerations
are necessary:—
That there be sufficient bearing surface for the arm to rest on.
That the box be of a convenient shape for insertion in the wheel.
That as large a body of oil as possible be kept in actual contact
with the arm by washing up as the wheel revolves.
That the column of oil may be in no case above the horizontal
level of the leakage point while the wheel is at rest.
Welding Steel Axles.
Many axles are now made of Bessemer steel. Generally speaking

this is neither more nor less than iron, the pores of which are filled up
with carbon or charcoal. The higher the steel the more carbon it
contains. If steel be heated it loses a portion of this carbon, and the
more it is heated the more it approaches its original state, viz. iron.
The welding of steel axles is said to be considerably assisted by
the use of iron filings and borax. This is only true in case the steel
should be over-heated, and even then only in degree.
Borax by itself is a very useful adjunct to this process, and it
should have a small quantity of sal-ammoniac added, to assist its
fusion or melting. The furnace or fire, which is to be used for the
welding process, should be clean and free from new coal, to prevent
sulphur getting on the steel. Of course, all coal has more or less
sulphur in it; but iron or steel cannot be successfully welded when
there is much sulphur in the fire, so it is well to be as careful in this
respect as possible.
Place the ends of the axles in a clean bright fire, heat to a bright
red heat, take them out, lap them over each other, and give them a
few smart blows with the sledge. Now well cover them with
powdered borax, and again put them into the fire and cover them up
with coked coal, give a strong even blast, and carefully watch the
appearance of the steel as the heat penetrates it, and see that all
parts of the weld are equally well heated. When the heat is raised as
high as the steel will safely bear (this knowledge can only be gained
by experience, so no rule can be given for ascertaining the degree of
heat, as it varies with the quality of the steel) take them out. Have
two men ready to use the sledges. Place the axles on the anvil,
securing them to prevent their slipping, and while one man places
his hammer full on the weld, give the extremity of the lap or weld a
smart blow or two, and if it adheres then both sledges can be applied
until a true and workmanlike weld is formed.
It sometimes happens that when the axles are heated ready for
welding and lapped, a light or a heavy blow, instead of uniting the
laps, only jars them apart. This is a sure sign that they have been
over-heated, and in this case it will be very difficult to form a weld at
all. The only way of getting over this difficulty is to heat it to as high a
degree as necessary, and put it in a vice and screw it up; the
surfaces will adhere in this way when the other means fail.
Another cause of failure is the too free use of borax. If too much is
used, it melts and runs about in the fire, unites with the dirt, and
generally blocks up the nozzle of the blast, causing a great deal of
trouble to dislodge. If the blast is not sufficient, then less heat is
generated than is necessary, and it is impossible to form a good
weld unless sufficient heat is applied.
Steel axles do not find great favour with the trade, although a large
quantity of them are used. They are unreliable, breaking and
fracturing without a moment’s warning, whereas an axle of faggoted
iron would only twist under the same circumstances, and could
easily be re-forged and set right again.
Setting Axles.
Setting axles is giving them the bend and slope required, in order
to fall in with the principles of the dished wheel. It is chiefly applied to
the axle-arm, and this is the most important part, setting the beds
being mere caprice.
The great object to be obtained is, to give the arm the right pitch
every way, to make the carriage run easy and as light as possible,
even in the absence of a plumb spoke. All carriages do not look best,
when running, with the bottom spoke plumb or vertical. In some of
the heavier coaches or carriages more slope or “pitch” has to be
given to the arm to carry the wheel away from the body, so as to
bring them to some specified track, in order to suit some particular
customer, so that we must be governed by circumstances.
There is a patent “axle-set,” but it is not of much assistance, for
half the smiths know nothing about it, and if they did it would not be
generally used, as the advantages derived from its use are not equal
to the trouble of using it. Besides, the wheels are not always dished
exactly alike, and it would require adjusting to each variety of wheel;
and again, the wheels are not always (though they ought to be)
ready; and when the smith knows the sort of vehicle he is working
upon he can give his axles the required pitch, within half a degree or
so, and the patent axle-set is, unfortunately, not capable of being
adjusted to an idea.
Fig. 21.
Fig. 21 shows a contrivance for setting the axles when cold, and
consists of an iron bar a, 2 feet 1 inch long, and about 2 inches
square at the fulcrum b. A hole is punched through the end to allow
the screw c to go through; this hole to be oval, to allow the screw to
move either way. At the end of this screw is an eye of sufficient size
to go on to the axle-arm. In setting the axle the eye is slipped on to
about the centre of the arm; the clevis, d, is placed on the bar a,
near the end; the fulcrum, b, is placed at the shoulder, either on top
or underneath, according as the axle may be required to set in or
out. When the fulcrum is laid on top, a strip of harness leather should
be placed on the axle bed, and on that, an iron e, of the shape of the
axle bed, and on the end of this the fulcrum is placed; then by
turning the screw the axle may be bent or set to any required pitch.
Fig. 22.
Fig. 23.
The figure shows the two ways of doing this, one with the bar or
lever on top and the other with the lever below.
Figs. 22 and 23 show two improved forms of axles.
Fig. 24.
Fig. 24 shows another variety of the axle-set. It consists of a bar
hooked on to the axletree in two places. The bar is fastened by the
clamp m, and fulcrum block f. The eyebolt, l, is hooked over the end
of the spindle or arm, and the adjustment of the latter is
accomplished by the screw, s, and the nuts j, k.
Weight of Round Iron per Foot.

Diameter. Diameter.
lbs. lbs.
Inch. Inch.
¼ ·163 2⅜ 14·7
⅜ ·368 2½ 16·3
½ ·654 2⅝ 18·0
⅝ 1·02 2¾ 19·7
¾ 1·47 2⅞ 21·6
⅞ 2·00 3 23·5
1 2·61 3⅛ 25·5
1⅛ 3·31 3¼ 27·6
1¼ 4·09 3⅜ 29·8
1⅜ 4·94 3½ 32·0
1½ 5·89 3⅝ 34·4
1⅝ 6·91 3¾ 36·8
1¾ 8·01 4 41·8
1⅞ 9·20 4¼ 47·2
2 10·4 4½ 53·0
2⅛ 11·8 5 65·4
2¼ 13·2
Weight of Square Iron per Foot.

Side of Square. Side of Square.
lbs. lbs.
Inch. Inch.
¼ ·208 2⅜ 18·8
⅜ ·468 2½ 20·8
½ ·833 2⅝ 22·9
⅝ 1·30 2¾ 25·2
¾ 1·87 2⅞ 27·5
⅞ 2·55 3 30·0
1 3·33 3⅛ 32·5
1⅛ 4·21 3¼ 35·2
1¼ 5·20 3⅜ 37·9
1⅜ 6·30 3½ 40·3
1½ 7·50 3⅝ 43·8
1⅝ 8·80 3¾ 46·8
1¾ 10·2 4 53·3
1⅞ 11·7 4¼ 60·2
2 13·3 4½ 67·5
2⅛ 15·0 5 83·3
2¼ 16·8
CHAPTER VIII.
SPRINGS.
Springs in locomotive vehicles are the elastic substances interposed
between the wheels and the load or passengers in order to intercept
the concussion caused by running over an uneven road, or in
meeting with any slight obstacle.
A great variety of substances have been used for this purpose,
such as leather, strips of hide, catgut, hempen cord, &c.; but these
have now been totally superseded by metal springs, so that what is
technically understood by the word “spring” is a plate or plates of
tempered steel properly shaped to play in any required mode.
It is very probable that the earliest steel springs were composed of
only one plate of metal. This was very defective in its action; and
unless it was restrained somewhat in the manner of the bow by the
string, it was liable to break on being subjected to a sharp
concussion.
There is no hard and fast rule by which the spring-maker can be
guided so as to proportion the strength and elasticity of his springs to
the load they are required to bear; and even were such a rule in
existence it would be practically useless, because the qualities of
spring steel differ so much that what is known in mathematics as a
“constant” could hardly be maintained. The only guide to the maker
in this respect is observation of the working of certain springs under
given loads, such springs being made of a certain quality of steel,
and any peculiar features that appear should be carefully noted
down for future reference and application.
Springs are of two kinds, single and double; i.e. springs tapering in
one direction from end to end, and those which taper in two opposite
directions from a common centre, as in the ordinary elliptic spring.
The process of making a spring is conducted in the following
manner:—
The longest or back plate being cut to the proper length, is
hammered down slightly at the extremities, and then curled round a
mandrel the size of the suspension bolt. The side of the plate which
is to fit against the others is then hollowed out by hammering; this is
called “middling.” The next plate is then cut rather shorter than the
first; the ends are tapered down so as not to disturb the harmony of
the curve. This plate is middled on both sides. A slit is then cut at
each end about ¾ of an inch in length and ⅜ inch wide, in which a
rivet head slides to connect it with the first plate, so that in whatever
direction the force acts these two plates sustain each other. At a little
distance from this rivet a stud is formed upon the under surface by a
punch, which forces out a protuberance which slides in a slit in the
next plate. The next plate goes through precisely the same
operations, except that it is 3 or 4 inches shorter at each end, and so
on with as many plates as the spring is to consist of. The last plate,
like the first, is of course only middled on one side.
The plates of which the spring is to be composed having thus
been prepared, have next to undergo the process of “hardening” and
“tempering.” This is a very important branch of the business, and will
bear a detailed description. There is no kind of tempering which
requires so much care in manipulation as that of springs. It is
necessary that the plates be carefully forged, not over-heated, and
not hammered too cold; one is equally detrimental with the other. To
guard against a plate warping in tempering, it is requisite that both
sides of the forging shall be equally well wrought upon with the
hammer; if not, the plates will warp and twist by reason of the
compression on one side being greater than on the other.[1]
The forge should be perfectly clean, and a good clean charcoal
fire should be used. Or if coal be used it must be burned to coke in
order to get rid of the sulphur, which would destroy the “life” of the
steel. Carefully insert the steel in the fire, and slowly heat it evenly
throughout its entire length; when the colour shows a light red,
plunge it into lukewarm water—cold water chills the outer surface too
rapidly—and let it lie in the water a short time. Animal oil is better
than water; either whale or lard oil is the best, or lard can be used
with advantage. The advantage of using oil is that it does not chill the
steel so suddenly, and there is less liability to crack it. This process
is called “hardening.”
Remove the hardened spring-plate from the water or oil and
prepare to temper it. To do this make a brisk fire with plenty of live
coals; smear the hardened plate with tallow, and hold it over the
coals, but do not urge the draught of the fire with the bellows while
so doing; let the fire heat the steel very gradually and evenly. If the
plate is a long one, move it slowly over the fire so as to receive the
heat equally. In a few moments the tallow will melt, then take fire,
and blaze for some time; while the blaze continues incline the plate,
or carefully incline or elevate either extremity, so that the blaze will
circulate from end to end and completely envelop it. When the flame
has died out, smear again with tallow and blaze it off as before. If the
spring is to undergo hard work the plates may be blazed off a third
time. Then let them cool themselves off upon a corner of the forge;
though they are often cooled by immersion in water, still it is not so
safe as letting them cool by themselves.
After tempering the spring-plates are “set,” which consists in any
warps or bumps received in the foregoing processes being put
straight by blows from a hammer. Care should be taken to have the
plates slightly warm while doing this to avoid fracturing or breaking
the plates.
The plates are now filed on all parts exposed to view, i.e. the
edges and points of the middle plates, the top and edges of the back
plate, and the top and edges of the shortest plate. They are then put
together and a rivet put through the spring at the point of greatest
thickness, and this holds, with the help of the studs before
mentioned, the plates together.
It is evident from the above description of a common mode of
making springs, that the operation is not quite so perfect as it might
be. The plates, instead of being merely tapered at the ends, ought to
be done so from the rivet to the points. And another thing, it would
surely make a better job of it if the plates were to bear their whole
width one on the other; in the middled plates they only get a bearing
on the edges, and the rain and dust will inevitably work into the
hollows in the plates, and it will soon form a magazine of rust, and
we all know what an affinity exists between iron and oxygen and the
result of it; as far as carriage springs are concerned, it very soon
destroys their elasticity and renders them useless and dangerous.
To prevent oxidation some makers paint the inner faces of the
springs, and this is in a measure successful, but the play of the
spring-plates one upon the other is sure to rub off some portions of
the paint, and we are just as badly off as ever. A far better plan
would be to cleanse the surfaces by means of acid, and then tin
them all over, and this would not be very expensive, and certainly
protect the plates of the spring longer than anything else.
The spiral springs, used to give elasticity to the seats, &c., are
tempered by heating them in a close vessel with bone dust or animal
charcoal, and, when thoroughly heated, cooled in a bath of oil. They
are tempered by putting them into an iron pan with tallow or oil, and
shaking them about over a brisk fire. The tallow will soon blaze, and
keeping them on the move will cause them to heat evenly. The steel
springs for fire-arms are tempered in this way, and are literally “fried
in oil.” If a long slender spring is needed with a low temper, it can be
made by simply beating the soft forging on a smooth anvil with a
smooth-faced hammer.
Setting and Tempering old Springs.
In setting up old springs where they are inclined to settle, first take
the longest plate (having separated all the plates) and bring it into
shape; then heat it for about 2 feet in the centre to a cherry red, and
cool it off in cold water as quick as possible. This will give the steel
such a degree of hardness that it will be liable to break if dropped on
the floor. To draw the temper hold it over the blaze, carrying
backward and forward through the fire until it is so hot that it will
sparkle when the hammer is drawn across it, and then cool off.
Another mode is to harden the steel, as before stated, and draw
the temper with oil or tallow—tallow is the best. Take a candle, carry
the spring as before through the fire, and occasionally draw the
candle over the length hardened, until the tallow will burn off in a
blaze, and then cool. Each plate is served in the same way.
Varieties of Springs.
The names given to springs are numerous, but the simple forms
are few, the greater part of the varieties being combinations of the
simple forms.
Fig. 25. Fig. 26.

The simple forms are the elliptic spring, the straight spring, and the
regular curve or C spring (Fig. 25). There are also one or two forms
of spring which have become obsolete. Such are the whip spring
(Fig. 26), and the reverse curved spring, which was superseded by
the last.
The elliptic spring is the one most commonly used at the present
day. Fig. 27, b, shows two of these united at the extremities by
means of a bolt; this is called a double elliptic spring. The elliptic
spring is sometimes used single in what are called under-spring
carriages, where the spring rests on the axle and is connected with
the framework of the body with an imitation spring or dumb iron to
complete the ellipse. Its technical name is an “under-spring.”
When four pairs of these springs are hinged together so as to form
four ellipses they constitute a set, and are used in carriages without
perches. Their technical name is “nutcracker spring.”
The straight springs are used in phaetons and tilburies, and are
called “single-elbow springs.”
The double straight spring is used in omnibuses, carts, &c., where
it is fixed across the angle at right angles. It is called a “double-elbow
spring.”
The regular curved spring is in form generally two-thirds of a circle,
one end of which is lengthened out into a tangent, which serves as a
base to fix it by in an upright position; the body is suspended from
the other extremity by means of leathern braces. Its general figure
has caused it to acquire the technical name of C spring. (See Fig.
25.)
The combination known as “telegraph spring” consists of eight
straight springs, when used for a four-wheeled carriage, and four
springs for a two-wheeled carriage. The Stanhope is suspended on
four of these springs. Two springs are fixed longitudinally on the
framework, and two transverse ones are suspended from these by
shackles, and on these latter the weight rests. They will bear a great
weight, and the body has the advantage of being placed two
removes from the concussion.
Fig. 27.
Fig. 27 shows some varieties of springs.
a Has semi-elliptical springs, hung upon the ends of C springs
attached to the axles.
b Has the usual elliptical springs between the bolster and axle.
c Has elastic wooden springs, which connect the axles and
support the beds.
d Has some elliptical springs, which also couple the axles a and b.
e Has a bolster hung upon C springs.
f Is a system of curved springs, with three points of connection to
the bed and two to the axles.
Weight of Elliptic Springs.
1¼ × 3 × 36 inch, weight about 28 lbs. per pair.
1¼ × 4 × 36 „ „ 34 „ „
1¼ × 4 × 38 „ „ 36 „ „
1½ × 3 × 36 „ „ 37 „ „
1½ × 4 × 36 „ „ 41 „ „
1½ × 4 × 38 „ „ 45 „ „
1½ × 5 × 36 „ „ 48 „ „
1½ × 5 × 38 „ „ 51 „ „
1½ × 5 × 40 „ „ 54 „ „
1¾ × 4 × 36 „ „ 49 „ „
1¾ × 4 × 38 „ „ 52 „ „
1¾ × 4 × 40 „ „ 55 „ „
1¾ × 5 × 36 „ „ 56 „ „
1¾ × 5 × 38 „ „ 60 „ „
1¾ × 5 × 40 „ „ 64 „ „
1¾ × 6 × 36 „ „ 64 „ „
1¾ × 6 × 38 „ „ 68 „ „
1¾ × 6 × 40 „ „ 73 „ „
2 × 4 × 36 „ „ 58 „ „
2 × 4 × 38 „ „ 62 „ „
2 × 4 × 40 „ „ 65 „ „
2 × 5 × 36 „ „ 63 „ „
2 × 5 × 38 „ „ 67 „ „
2 × 5 × 40 „ „ 72 „ „
2 × 6 × 36 „ „ 75 „ „
2 × 6 × 38 „ „ 78 „ „
2 × 6 × 40 „ „ 85 „ „
FOOTNOTES:
[1] It is the plates that are tempered and hardened, not the
spring.
CHAPTER IX.
WHEEL-PLATES AND FORE-CARRIAGES.

The following is given in the “Coachmaker’s Handbook” under the
heading of “Short and Easy Turning:”—
“To bring a carriage into a different course from a straight one
requires a circular motion, and at half a turn a carriage has
established itself in a right angle to its position when at rest.
“A two-wheeled vehicle turns on one wheel, which forms the
centre at the place where it touches the ground, and the opposite
wheel forms the circle struck from the said centre. The body in this
instance follows the circular motion exactly as the axle, and
consequently maintains a steady position above the wheels.
“A four-wheeled vehicle remains in a straight line when first the
front pair of wheels are turned under, then by the effect of the
draught the hind pair of wheels follow in a wider circle. To effect a
turning we bring the front axle first in a corresponding direction with
the desired turn.
“We make distinction between the moment of turning, or the
angular position of the axles previous to the turning itself, and the
effected turning of a vehicle round a centre or king bolt, according to
the construction of the carriage part. The wheels have to be brought
in a position corresponding with the direction of the turning. The
body must be fully supported after the turning, and the front or dickey
of a carriage must stand in a right angle to fore axletree.
“We have to consider a few points relative to the height of the front
wheel, and the elevation of the body above the ground, which
averages 30 inches. To give a front wheel its proper height (between
3 feet 4 inches and 3 feet 6 inches) and have it turn a full circle, we
sweep the body at the required place, viz. put in the wheel house of
a proportioned length, and a depth between 3 and 4½ inches.
“The front carriage part is fastened round the king bolt, turning that
part horizontal. This action causes the front wheel to describe a
circle, whose diameter is the width of the track; but as the wheel
leans over at the top through the dish, we have a larger circle in the
middle and top of the wheel. We, therefore, first find a top circle,
having a diameter equal to the width between the highest point of the
wheel, and a side circle following the termination of the cross
diameter of the wheel, having as a centre the king bolt.”
In Fig. 28, which is drawn to
quarter-inch scale, the
horizontal line a is the axletree,
b is the wheel at rest, c is the
wheel on full lock, d is the back
of the arch, e shows the circle
that the wheel will describe on
the ground as it moves
backward, and f is the circle the
back of the wheel will describe
in the air at the same
movement. It will be seen by
Fig. 28.
this that when the wheel is on
half lock the back part of it will
come in contact with the arch, and that when on full lock it will have
travelled right away from it. It follows, therefore, that if we want to
find out the right position for the perch bolt to occupy, we must not
measure the circle the wheel will describe on the ground, but the one
described in the air. We must, therefore, measure along the line f,
and carry that measurement along to d.
The length of the line a to d is exactly 3 feet. Now the position of
the perch bolt, or centre point on which the wheels lock round, need
not be, and in fact very rarely is, in the same vertical line as the
axletree. By compassing the beds or timbers on which the fore part

Download pdf Advances In Self Organizing Maps Learning Vector Quantization Clustering And Data Visualization Proceedings Of The 13Th International Workshop Wsom 2019 Barcelona Spain June 26 28 2019 Alfred ebook full chapter

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Download pdf Advances In Self Organizing Maps Learning Vector Quantization Clustering And Data Visualization Proceedings Of The 13Th International Workshop Wsom 2019 Barcelona Spain June 26 28 2019 Alfred ebook full chapter

Uploaded by

Copyright:

Available Formats

Advances in Self-Organizing Maps,

Learning Vector Quantization,

Advances in Self-Organizing Maps, Learning Vector

Advances in Self Organizing Maps and Learning Vector

Business Information Systems: 22nd International

Business Information Systems: 22nd International

Graph Drawing and Network Visualization 26th

Swarm Intelligence 12th International Conference ANTS

Big Data 7th CCF Conference BigData 2019 Wuhan China

Evolving OpenMP for Evolving Architectures 14th

** Indexing: The books of this series are submitted to ISI Proceedings,

More information about this series at http://www.springer.com/series/11156

Cecilio Angulo José David Martín Guerrero

ISSN 2194-5357 ISSN 2194-5365 (electronic)

to the international researchers who agreed to participate as plenary speakers in

Teuvo Kohonen (Honorary Chairman), Finland

Cecilio Angulo Universitat Politècnica de Catalunya, Spain

Alexander Gepperth University of Applied Sciences, HAW Fulda,

Self-organizing Maps: Theoretical Developments

Practical Applications of Self-Organizing Maps, Learning Vector

Using SOM-Based Visualization to Analyze the Financial

Learning Vector Quantization: Theoretical Developments

Prototype-Based Classiﬁers in the Presence of Concept Drift:

Theoretical Developments in Clustering, Deep Learning

Life Science Applications

Searching for the Origins of Life – Detecting RNA Life Signatures

Jérémy Fix(B) and Hervé Frezza-Buet

LORIA, CNRS, CentraleSupélec, Université Paris-Saclay, 57000 Metz, France

Abstract. This paper introduces representations and measurements for

Keywords: Recurrent self-organizing map · Sequence processing ·

Let us consider a stream of inputs ξ t ∈ X available at each successive time step

Once i is determined consecutively to the submission of ξ t , the prototypes in

Algorithm 1. Architecture update at time t.

The representations presented so far enables to unveil the inner dynamics of

3.1 Ambiguous Observations

observations. There are 8 states which provide an observation that is ambiguous

0.4 0.4 0.4

1.0 1.0 1.0

0.8 0.8 0.8

Fig. 1. Observations from ABCDEFEDCB. The protographs are recorded at t =

3.2 Long Term Dependencies

protograph around position 0.3). This is the simplest to be identiﬁed because it

0.4 0.4 0.4

1.0 1.0 1.0

0.8 0.8 0.8

Fig. 2. Observations from AAAAAAAF. The protographs are recorded at t =

3.3 Adapting to a Changing Dynamics

0.2 0.2 0.5

1.0 1.0 0.3

3.4 Noisy Observations

We now perform an experiment to test the robustness of the algorithm in

3.5 Perturbed by a Noise State

ABCDEFEDCB. The clean sequence of observations from ABCDEFEDCB is therefore

Acknowledgement. This work is supported by the European Interreg Grande

Xiaofeng Ma, Michael Kirby(B) , and Chris Peterson

Colorado State University, Fort Collins, CO 80523, USA

Abstract. A flag is a nested sequence of vector spaces. The type of the

Keywords: Self-Organizing Mappings · SOM · Flag manifolds ·

Self-Organizing Mappings (SOMs) were introduced as a means to see data in

c Springer Nature Switzerland AG 2020

Motivated by the subspace approach to data analytics we proposed a version

2 Introduction to Flag Manifold with Data Analysis

Let us ﬁrst introduce the ﬂag manifold. A f lag of type (n1 , n2 , . . . , nd ; n) is a

Fig. 1. Illustration of a nested sequence of subspaces corresponding to a point on the