Download Full Version ebook - Visit ebooknice.

com

(Ebook) Deep Learning Applications, Volume 2 by M.

Arif Wani, Taghi M. Khoshgoftaar, Vasile Palade
ISBN 9789811567582, 9789811567599, 9811567581,
981156759X
https://ebooknice.com/product/deep-learning-applications-
volume-2-22476784

Click the button below to download

DOWLOAD EBOOK

Discover More Ebook - Explore Now at ebooknice.com

 Instant digital products (PDF, ePub, MOBI) ready for you
Download now and discover formats that fit your needs...

Start reading on any device today!

(Ebook) Biota Grow 2C gather 2C cook by Loucas, Jason;

Viles, James ISBN 9781459699816, 9781743365571,
9781925268492, 1459699815, 1743365578, 1925268497
https://ebooknice.com/product/biota-grow-2c-gather-2c-cook-6661374

ebooknice.com

(Ebook) Matematik 5000+ Kurs 2c Lärobok by Lena

Alfredsson, Hans Heikne, Sanna Bodemyr ISBN 9789127456600,
9127456609
https://ebooknice.com/product/matematik-5000-kurs-2c-larobok-23848312

ebooknice.com

(Ebook) SAT II Success MATH 1C and 2C 2002 (Peterson's SAT

II Success) by Peterson's ISBN 9780768906677, 0768906679

https://ebooknice.com/product/sat-ii-success-
math-1c-and-2c-2002-peterson-s-sat-ii-success-1722018

ebooknice.com

(Ebook) Master SAT II Math 1c and 2c 4th ed (Arco Master

the SAT Subject Test: Math Levels 1 & 2) by Arco ISBN
9780768923049, 0768923042
https://ebooknice.com/product/master-sat-ii-math-1c-and-2c-4th-ed-
arco-master-the-sat-subject-test-math-levels-1-2-2326094

ebooknice.com
(Ebook) Cambridge IGCSE and O Level History Workbook 2C -
Depth Study: the United States, 1919-41 2nd Edition by
Benjamin Harrison ISBN 9781398375147, 9781398375048,
1398375144, 1398375047
https://ebooknice.com/product/cambridge-igcse-and-o-level-history-
workbook-2c-depth-study-the-united-states-1919-41-2nd-edition-53538044

ebooknice.com

(Ebook) Deep Learning Applications, Volume 3 (Advances in

Intelligent Systems and Computing) by M. Arif Wani &
Bhiksha Raj & Feng Luo & Dejing Dou; Editors ISBN
9789811633560, 9811633568
https://ebooknice.com/product/deep-learning-applications-
volume-3-advances-in-intelligent-systems-and-computing-36372730

ebooknice.com

(Ebook) Generative Adversarial Learning: Architectures and

Applications by Roozbeh Razavi-Far, Ariel Ruiz-
GarciaVasile, Vasile Palade, Juergen Schmidhuber ISBN
9783030913908, 3030913902
https://ebooknice.com/product/generative-adversarial-learning-
architectures-and-applications-38342882

ebooknice.com

(Ebook) Introduction to Deep Learning for Engineers: Using

Python and Google Cloud Platform by Tariq M. Arif ISBN
9781681739144, 1681739143
https://ebooknice.com/product/introduction-to-deep-learning-for-
engineers-using-python-and-google-cloud-platform-33289320

ebooknice.com

(Ebook) Advances in Hybridization of Intelligent Methods:

Models, Systems and Applications by Ioannis
Hatzilygeroudis, Vasile Palade (eds.) ISBN 9783319667898,
9783319667904, 3319667890, 3319667904
https://ebooknice.com/product/advances-in-hybridization-of-
intelligent-methods-models-systems-and-applications-6790406

ebooknice.com
Advances in Intelligent Systems and Computing 1232

M. Arif Wani
Taghi M. Khoshgoftaar
Vasile Palade Editors

Deep Learning
Applications,
Volume 2
Advances in Intelligent Systems and Computing

Volume 1232

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, School of Computer Science and 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

M. Arif Wani Taghi M. Khoshgoftaar
• •

Vasile Palade
Editors

Deep Learning Applications,

Volume 2

123
Editors
M. Arif Wani Taghi M. Khoshgoftaar
Department of Computer Science Computer and Electrical Engineering
University of Kashmir Florida Atlantic University
Srinagar, India Boca Raton, FL, USA

Vasile Palade
Faculty of Engineering and Computing
Coventry University
Coventry, UK

ISSN 2194-5357 ISSN 2194-5365 (electronic)

Advances in Intelligent Systems and Computing
ISBN 978-981-15-6758-2 ISBN 978-981-15-6759-9 (eBook)
https://doi.org/10.1007/978-981-15-6759-9

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature
Singapore Pte Ltd. 2021
This work is subject to copyright. All rights are solely and exclusively licensed 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 Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
Preface

Machine learning algorithms have influenced many aspects of our day-to-day living
and transformed major industries around the world. Fueled by an exponential
growth of data, improvements in computer hardware, scalable cloud resources, and
accessible open-source frameworks, machine learning technology is being used by
companies in big and small alike for innumerable applications. At home, machine
learning models are suggesting TV shows, movies, and music for entertainment,
providing personalized ecommerce suggestions, shaping our digital social net-
works, and improving the efficiency of our appliances. At work, these data-driven
methods are filtering our emails, forecasting trends in productivity and sales, tar-
geting customers with advertisements, improving the quality of video conferences,
and guiding critical decisions. At the frontier of machine learning innovation are
deep learning systems, a class of multi-layered networks is capable of automatically
learning meaningful hierarchical representations from a variety of structured and
unstructured data. Breakthroughs in deep learning allow us to generate new rep-
resentations, extract knowledge, and draw inferences from raw images, video
streams, text and speech, time series, and other complex data types. These powerful
deep learning methods are being applied to new and exciting real-world problems in
medical diagnostics, factory automation, public safety, environmental sciences,
autonomous transportation, military applications, and much more.
The family of deep learning architectures continues to grow as new methods and
techniques are developed to address a wide variety of problems. A deep learning
network is composed of multiple layers that form universal approximators capable
of learning any function. For example, the convolutional layers in Convolutional
Neural Networks use shared weights and spatial invariance to efficiently learn
hierarchical representations from images, natural language, and temporal data.
Recurrent Neural Networks use backpropagation through time to learn from vari-
able length sequential data. Long Short-Term Memory networks are a type of
recurrent network capable of learning order dependence in sequence prediction
problems. Deep Belief Networks, Autoencoders, and other unsupervised models
generate meaningful latent features for downstream tasks and model the underlying
concepts of distributions by reconstructing their inputs. Generative Adversarial

v
vi Preface

Networks simultaneously learn generative models capable of producing new data

from distribution and discriminative models that can distinguish between real and
artificial images. Transformer Networks combine encoders and decoders with
attention layers for improved sequence-to-sequence learning. Network architecture
search automates the designs of these deep models by optimizing performance over
the hyperparameter space. As a result of these advances, and many others, deep
learning is revolutionizing complex problem domains with state-of-the-art results
and, in some cases, is a way superior to the human performances.
This book explores some of the latest applications in deep learning and includes
a variety of architectures and novel deep learning techniques. Deep models are
trained to recommend products, diagnose medical conditions or faults in industrial
machines, detect when a human falls, and recognize solar panels in aerial images.
Sequence models are used to capture driving behaviors and identify radio trans-
mitters from temporal data. Residual networks are used to detect human targets in
indoor environments, algorithm incorporating thresholding strategy is used to
identify fraud within highly imbalanced data, and hybrid methods are used to locate
vehicles during satellite outages. Multi-adversarial variational autoencoder network
is used for image synthesis and classification and finally parameter continuation
method is used for non-convex optimization of deep neural networks. We believe
that these recent deep learning methods and applications illustrated in this book
capture some of the most exciting advances in deep learning.

Srinagar, India M. Arif Wani

Boca Raton, USA Taghi M. Khoshgoftaar
Coventry, UK Vasile Palade
Contents

Deep Learning-Based Recommender Systems . . . . . . . . . . . . . . . . . . . . 1

Meshal Alfarhood and Jianlin Cheng
A Comprehensive Set of Novel Residual Blocks for Deep Learning
Architectures for Diagnosis of Retinal Diseases from Optical
Coherence Tomography Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Sharif Amit Kamran, Sourajit Saha, Ali Shihab Sabbir,
and Alireza Tavakkoli
Three-Stream Convolutional Neural Network for Human
Fall Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
Guilherme Vieira Leite, Gabriel Pellegrino da Silva, and Helio Pedrini
Diagnosis of Bearing Faults in Electrical Machines Using Long
Short-Term Memory (LSTM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
Russell Sabir, Daniele Rosato, Sven Hartmann, and Clemens Gühmann
Automatic Solar Panel Detection from High-Resolution Orthoimagery
Using Deep Learning Segmentation Networks . . . . . . . . . . . . . . . . . . . . 101
Tahir Mujtaba and M. Arif Wani
Training Deep Learning Sequence Models to Understand
Driver Behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
Shokoufeh Monjezi Kouchak and Ashraf Gaffar
Exploiting Spatio-Temporal Correlation in RF Data
Using Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
Debashri Roy, Tathagata Mukherjee, and Eduardo Pasiliao
Human Target Detection and Localization with Radars
Using Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
Michael Stephan, Avik Santra, and Georg Fischer

vii
viii Contents

Thresholding Strategies for Deep Learning with Highly Imbalanced

Big Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
Justin M. Johnson and Taghi M. Khoshgoftaar
Vehicular Localisation at High and Low Estimation Rates During
GNSS Outages: A Deep Learning Approach . . . . . . . . . . . . . . . . . . . . . 229
Uche Onyekpe, Stratis Kanarachos, Vasile Palade,
and Stavros-Richard G. Christopoulos
Multi-Adversarial Variational Autoencoder Nets for Simultaneous
Image Generation and Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
Abdullah-Al-Zubaer Imran and Demetri Terzopoulos
Non-convex Optimization Using Parameter Continuation Methods
for Deep Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
Harsh Nilesh Pathak and Randy Clinton Paffenroth

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299

Editors and Contributors

About the Editors

Dr. M. Arif Wani is a Professor at the University of Kashmir, having previously

served as a Professor at California State University, Bakersfield. He completed his
M.Tech. in Computer Technology at the Indian Institute of Technology, Delhi, and
his Ph.D. in Computer Vision at Cardiff University, UK. His research interests are
in the area of machine learning, with a focus on neural networks, deep learning,
inductive learning, and support vector machines, and with application to areas that
include computer vision, pattern recognition, classification, prediction, and analysis
of gene expression datasets. He has published many papers in reputed journals and
conferences in these areas. Dr. Wani has co-authored the book ‘Advances in Deep
Learning,’ co-edited the book ‘Deep Learning Applications,’ and co-edited 17
conference proceeding books in machine learning and applications area. He is a
member of many academic and professional bodies, e.g., the Indian Society for
Technical Education, Computer Society of India, and IEEE USA.

Dr. Taghi M. Khoshgoftaar is the Motorola Endowed Chair professor of the

Department of computer and electrical engineering and Computer Science, Florida
Atlantic University, and the Director of NSF Big Data Training and Research
Laboratory. His research interests are in big data analytics, data mining and
machine learning, health informatics and bioinformatics, social network mining,
and software engineering. He has published more than 750 refereed journal and
conference papers in these areas. He was the Conference Chair of the IEEE
International Conference on Machine Learning and Applications (ICMLA 2019).
He is the Co-Editor-in-Chief of the Journal of Big Data. He has served on orga-
nizing and technical program committees of various international conferences,
symposia, and workshops. He has been a Keynote Speaker at multiple international

ix
x Editors and Contributors

conferences and has given many invited talks at various venues. Also, he has served
as North American Editor of the Software Quality Journal, was on the editorial
boards of the journals Multimedia Tools and Applications, Knowledge and
Information Systems, and Empirical Software Engineering, and is on the editorial
boards of the journals Software Quality, Software Engineering and Knowledge
Engineering, and Social Network Analysis and Mining.

Dr. Vasile Palade is currently a Professor of Artificial Intelligence and Data

Science at Coventry University, UK. He previously held several academic and
research positions at the University of Oxford—UK, University of Hull—UK, and
the University of Galati—Romania. His research interests are in the area of machine
learning, with a focus on neural networks and deep learning, and with main
application to image processing, social network data analysis and web mining,
smart cities, health, among others. Dr. Palade is author and co-author of more than
170 papers in journals and conference proceedings as well as several books on
machine learning and applications. He is an Associate Editor for several reputed
journals, such as Knowledge and Information Systems and Neurocomputing. He
has delivered keynote talks to international conferences on machine learning and
applications. Dr. Vasile Palade is an IEEE Senior Member.

Contributors

Meshal Alfarhood Department of Electrical Engineering and Computer Science,

University of Missouri-Columbia, Columbia, USA
M. Arif Wani Department of Computer Science, University of Kashmir, Srinagar,
India
Jianlin Cheng Department of Electrical Engineering and Computer Science,
University of Missouri-Columbia, Columbia, USA
Stavros-Richard G. Christopoulos Institute for Future Transport and Cities,
Coventry University, Coventry, UK;
Faculty of Engineering, Coventry University, Coventry, UK
Randy Clinton Paffenroth Worcester Polytechnic Institute, Mathematical
Sciences Computer Science & Data Science, Worcester, MA, USA
Gabriel Pellegrino da Silva Institute of Computing, University of Campinas,
Campinas, SP, Brazil
Georg Fischer Friedrich-Alexander-University Erlangen-Nuremberg, Erlangen,
Germany
Ashraf Gaffar Arizona State University, Tempe, USA
Editors and Contributors xi

Clemens Gühmann Chair of Electronic Measurement and Diagnostic Technology

& Technische Universität Berlin, Berlin, Germany
Sven Hartmann SEG Automotive Germany GmbH, Stuttgart, Germany
Abdullah-Al-Zubaer Imran University of California, Los Angeles, CA, USA
Justin M. Johnson Florida Atlantic University, Boca Raton, FL, USA
Sharif Amit Kamran University of Nevada, Reno, NV, USA
Stratis Kanarachos Faculty of Engineering, Coventry University, Coventry, UK
Taghi M. Khoshgoftaar Florida Atlantic University, Boca Raton, FL, USA
Shokoufeh Monjezi Kouchak Arizona State University, Tempe, USA
Guilherme Vieira Leite Institute of Computing, University of Campinas,
Campinas, SP, Brazil
Tahir Mujtaba Department of Computer Science, University of Kashmir,
Srinagar, India
Tathagata Mukherjee Computer Science, University of Alabama, Huntsville,
AL, USA
Harsh Nilesh Pathak Expedia Group, Seattle, WA, USA
Uche Onyekpe Institute for Future Transport and Cities, Coventry University,
Coventry, UK;
Research Center for Data Science, Coventry University, Coventry, UK
Vasile Palade Research Center for Data Science, Coventry University, Coventry,
UK
Eduardo Pasiliao Munitions Directorate, Air Force Research Laboratory, Eglin
AFB, Valparaiso, FL, USA
Helio Pedrini Institute of Computing, University of Campinas, Campinas, SP,
Brazil
Daniele Rosato SEG Automotive Germany GmbH, Stuttgart, Germany
Debashri Roy Computer Science, University of Central Florida, Orlando, FL,
USA
Russell Sabir SEG Automotive Germany GmbH, Stuttgart, Germany;
Chair of Electronic Measurement and Diagnostic Technology & Technische
Universität Berlin, Berlin, Germany
Ali Shihab Sabbir Center for Cognitive Skill Enhancement, Independent
University Bangladesh, Dhaka, Bangladesh
xii Editors and Contributors

Sourajit Saha Center for Cognitive Skill Enhancement, Independent University

Bangladesh, Dhaka, Bangladesh
Avik Santra Infineon Technologies AG, Neubiberg, Germany
Michael Stephan Infineon Technologies AG, Neubiberg, Germany;
Friedrich-Alexander-University Erlangen-Nuremberg, Erlangen, Germany
Alireza Tavakkoli University of Nevada, Reno, NV, USA
Demetri Terzopoulos University of California, Los Angeles, CA, USA
Deep Learning-Based Recommender
Systems

Meshal Alfarhood and Jianlin Cheng

Abstract The term “information overload” has gained popularity over the last few
years. It defines the difficulties people face in finding what they want from a huge
volume of available information. Recommender systems have been recognized to be
an effective solution to such issues, such that suggestions are made based on users’
preferences. This chapter introduces an application of deep learning techniques in
the domain of recommender systems. Generally, collaborative filtering approaches,
and Matrix Factorization (MF) techniques in particular, are widely known for their
convincing performance in recommender systems. We introduce a Collaborative
Attentive Autoencoder (CATA) that improves the matrix factorization performance
by leveraging an item’s contextual data. Specifically, CATA learns the proper features
from scientific articles through the attention mechanism that can capture the most
pertinent parts of information in order to make better recommendations. The learned
features are then incorporated into the learning process of MF. Comprehensive exper-
iments on three real-world datasets have shown our method performs better than other
state-of-the-art methods according to various evaluation metrics. The source code of
our model is available at: https://github.com/jianlin-cheng/CATA.

This chapter is an extended version of our published paper at the IEEE ICMLA conference 2019
[1]. This chapter incorporates new experimental contributions compared to the original confere-
nce paper.

M. Alfarhood (B) · J. Cheng

Department of Electrical Engineering and Computer Science,
University of Missouri-Columbia, Columbia, USA
e-mail: may82@missouri.edu
J. Cheng
e-mail: chengji@missouri.edu

© The Editor(s) (if applicable) and The Author(s), under exclusive license 1
to Springer Nature Singapore Pte Ltd. 2021
M. A. Wani et al. (eds.), Deep Learning Applications, Volume 2,
Advances in Intelligent Systems and Computing 1232,
https://doi.org/10.1007/978-981-15-6759-9_1
2 M. Alfarhood and J. Cheng

1 Introduction

The era of e-commerce has vastly changed people’s lifestyles during the first part
of the twenty-first century. People today tend to do many of their daily routines
online, such as shopping, reading the news, and watching movies. Nevertheless,
consumers often face difficulties while exploring related items such as new fashion
trends because they are not aware of their existence due to the overwhelming amount
of information available online. This phenomenon is widely known as “information
overload”. Therefore, Recommender Systems (RSs) are a critical solution for helping
users make decisions when there are lots of choices. RSs have been integrated into
and have become an essential part of every website due to their impact on increasing
customer interactions, attracting new customers, and growing businesses’ revenue.
Scientific article recommendation is a very common application for RSs. It keeps
researchers updated on recent related work in their field. One traditional way to
find relevant articles is to go through the references section in other articles. Yet,
this approach is biased toward heavily cited articles, such that new relevant articles
with higher impact have less chance to be found. Another method is to search for
articles using keywords. Although this technique is popular among researchers, they
must filter out a tremendous number of articles from the search results to retrieve
the most suitable articles. Moreover, all users get the same search results with the
same keywords, and these results are not personalized based on the users’ personal
interests. Thus, recommendation systems can address this issue and help scientists
and researchers find valuable articles while being aware of recent related work.
Over the last few decades, a lot of effort has been made by both academia and
industry on proposing new ideas and solutions for RSs, which ultimately help ser-
vice providers in adopting such models in their system architecture. The research in
RSs has evolved remarkably following the Netflix prize competition1 in 2006, where
the company offered one million dollars for any team that could improve their rec-
ommendation accuracy by 10%. Since that time, collaborative filtering models and
matrix factorization techniques in particular have become the most common models
due to their effective performance. Generally, recommendation models are classified
into three categories: Collaborative Filtering Models (CF), Content-Based Filter-
ing models (CBF), and hybrid models. CF models [2–4] focus on users’ histories,
such that users with similar past behaviors tend to have similar future tastes. On the
other hand, CBF models work by learning the item’s features from its informational
description, such that two items are possibly similar to each other if they share more
characteristics. For example, two songs are similar to each other if they both share
the same artist, genre, tempo, energy, etc. However, similarities between items in CF
models are different such that two items are likely similar to each other once they are
rated by multiple users in the same manner, even though those items have different
characteristics.

1 www.netflixprize.com.
Deep Learning-Based Recommender Systems 3

Generally, CF models function better than CBF models. However, CF performance

drops substantially when users or items have an insufficient amount of feedback
data. This problem is defined as the data sparsity problem. To tackle data sparseness,
hybrid models have been widely proposed in recent works [5–8], in which content
information, used in CBF models, is incorporated with CF models to improve the
system performance. Hybrid models are divided into two sub-categories according
to how models are trained: loosely coupled models and tightly coupled models [7].
Loosely coupled models train CF and CBF models separately, like ensembles, and
then determine the final score based on the scores of the two separated models. On
the other hand, the tightly coupled models train both CF and CBF models jointly.
In joint training, both models cooperate with one another to calculate the final score
under the same loss function.
Even though traditional recommendation approaches have achieved great success
over the last years, they still have shortcomings in accurately modeling complex
(e.g., non-linear) relationship between users and items. Alternatively, deep neural
networks are universal function approximators that are capable of modeling any con-
tinuous function. Recently, Deep Learning (DL) has become an effective approach
for most data mining problems. DL meets recommendation systems in the last few
years. One of the first works that applied DL concept for CF recommendations was
Restricted Boltzmann Machines (RBM) [4]. However, this approach was not deep
enough (two layers only) to learn users’ tastes from their histories, and it also did not
take contextual information into consideration. Recently, Collaborative Deep Learn-
ing (CDL) [7] has become a very popular deep learning technique in RSs due to its
promising performance. CDL can be viewed as an updated version of Collaborative
Topic Regression (CTR) [5] by substituting the Latent Dirichlet Allocation (LDA)
topic modeling with a Stacked Denoising Autoencoder (SDAE) to learn from item
contents, and then integrating the learned latent features into a Probabilistic Matrix
Factorization (PMF). Lately, Collaborative Variational Autoencoder (CVAE) [8] has
been proposed to learn deep item latent features via a variational autoencoder. The
authors show that their model learns better item features than CDL because their
model infers the latent variable distribution in latent space instead of observation
space. However, both CDL and CVAE models assume that all parts of their model’s
contribution are the same for their final predictions.
Hence, in this work, we propose a deep learning-based model named Collaborative
Attentive Autoencoder (CATA) for recommending scientific articles. In particular,
we integrate the attention mechanism into our unsupervised deep learning process
to identify an item’s features. We learn the item’s features from the article’s textual
information (e.g., the article’s title and abstract) to enhance the recommendation
quality. The compressed low-dimensional representation learned by the unsupervised
model is incorporated then into the matrix factorization approach for our ultimate
recommendation. To demonstrate the capability of our proposed model to generate
more relevant recommendations, we conduct inclusive experiments on three real-
world datasets, which are taken from the CiteULike2 website, to evaluate CATA

2 www.citeulike.org.
4 M. Alfarhood and J. Cheng

against multiple recent works. The experimental results prove that our model can
extract more constructive information from an article’s contextual data than other
models. More importantly, CATA performs very well where the data sparsity is
extremely high.
The remainder of this chapter is organized in the following manner. First, we
demonstrate the matrix factorization method in Sect. 2. We introduce our model,
CATA, in Sect. 3. The experimental results of our model against the state-of-the-art
models are discussed thoroughly in Sect. 4. We then conclude our work in Sect. 5.

2 Background

Our work is designed and evaluated on recommendations with implicit feedback.

Thus, in this section, we describe the well-known collaborative filtering approach,
Matrix Factorization, for implicit feedback problems.

2.1 Matrix Factorization

Matrix Factorization (MF) [2] is the most popular CF method, mainly due to its
simplicity and efficiency. The idea behind MF is to decompose the user-item matrix,
R ∈ Rn×m , into two lower dimensional matrices, U ∈ Rn×d and V ∈ Rm×d , such
that the inner product of U and V will approximate the original matrix R, where d
is the dimension of the latent factors, such that d  min(n, m). n and m correspond
to the number of users and items in the system. Figure 1 illustrates the MF process.

R ≈ U · VT (1)

MF optimizes the values of U and V by minimizing the sum of the squared

difference between the actual values and the model predictions with adding two
regularization terms, as shown here:
 Ii j λu λv  
L= (ri j − u i v Tj )2 + u i 2 + v j 2 (2)
i, j∈R
2 2 2

where Ii j is an indicator function that equals 1 if useri has rated item j , and 0 if
otherwise. Also, ||U || and ||V || are the Euclidean norms, and λu , λv are two regu-
larization terms preventing the values of U and V from being too large. This avoid
model overfitting.
Explicit data, such as ratings (ri j ) are not regularly available. Therefore, Weighted
Regularized Matrix Factorization (WRMF) [9] introduces two modifications to the
previous objective function to make it work for implicit feedback. The optimization
Deep Learning-Based Recommender Systems 5

Fig. 1 Matrix factorization illustration

process in this case runs through all user-item pairs with different confidence levels
assigned to each pair, as in the following:
 ci j λu λv  
L= ( pi j − u i v Tj )2 + u i 2 + v j 2 (3)
i, j∈R
2 2 2

where pi j is the user preference score with a value of 1 when useri and item j have
an interaction, and 0 otherwise. ci j is a confidence variable where its value shows
how confident the user like the item. In general, ci j = a when pi j = 1, and ci j = b
when pi j = 0, such that a > b > 0.
Stochastic Gradient Decent (SGD) [10] and Alternating Least Squares (ALS) [11]
are two optimization methods that can be used to minimize the objective function
of MF in Eq. 2. The first method, SGD, loops over each single training sample and
then computes the prediction error as ei j = ri j − u i v Tj . The gradient of the objective
function with respect to u i and v j can be computed as follows:

∂L 
=− Ii j (ri j − u i v Tj )v j + λu u i
∂u i j
 (4)
∂L
=− Ii j (ri j − u i v Tj )u i + λv v j
∂v j i

After calculating the gradient, SGD updates the user and item latent factors in the
opposite direction of the gradient using the following equations:
6 M. Alfarhood and J. Cheng
⎛ ⎞

ui ← ui + α ⎝ Ii j ei j v j − λu u i ⎠
j
 (5)

vj ← vj + α Ii j ei j u i − λ j v j
i

where α is the learning rate.

Even though SGD is easy to implement and generally faster than ALS in some
cases, it is not suitable to use with implicit feedback, since looping over each single
training sample is not practical. ALS works better in this case. ALS iteratively opti-
mizes U while V is fixed, and vice versa. This optimization process is repeated until
the model converges.
To determine what user and item vector values minimize the objective function
for implicit data (Eq. 3), we first take the derivative of L with respect to u i .

∂L 
=− ci j ( pi j − u i v Tj )v j + λu u i
∂u i j

0 = −Ci (Pi − u i V T )V + λu u i
0 = −Ci V Pi + Ci V u i V T + λu u i (6)
V Ci Pi = u i V Ci V + λu u i
T

V Ci Pi = u i (V Ci V T + λu I )
ui = V Ci Pi (V Ci V T + λu I )−1
ui = (V Ci V T + λu I )−1 V Ci Pi

where I is the identity matrix.

Similarly, taking the derivative of L with respect to v j leads to

v j = (U C j U T + λv I )−1 U C j P j (7)

3 Proposed Model

In this section, we illustrate our proposed model in depth. The intuition behind our
model is to learn the latent factors of items in PMF with the use of available side
textual contents. We use an attentive unsupervised model to catch more plentiful
information from the available data. The architecture of our model is displayed in
Fig. 2. We first define the problem with implicit feedback before we go through the
details of our model.
Deep Learning-Based Recommender Systems 7

λu λv

X̂ j

Decoder

Attention

Ui Vj Zj X Softmax

Rij Encoder

Xj
i = 1:n j = 1:m

Fig. 2 Collaborative attentive autoencoder architecture

3.1 Problem Definition

User-item interaction data is the primary source for training recommendation

engines. This data can be either collected in an explicit or implicit manner. In explicit
data, users directly express their opinion about an item using the rating system to
show how much they like that item. The user’s ratings usually vary from one-star to
five-stars with five being very interested and one being not interested. This type of
data is very useful and reliable due to the fact that it represents the actual feeling of
users about items. However, users’ ratings occasionally are not available due to the
difficulty of obtaining users’ explicit opinions. In this case, implicit feedback can be
obtained indirectly from the user’s behavior such as user clicks, bookmarks, or the
time spent viewing an item. For instance, if a user listens to a song 10 times in the
last two days, he or she most likely likes this song. Thus, implicit data is more preva-
lent and easier to collect, but it is generally less reliable than explicit data. Also, all
the observed interactions in implicit data constitute positive feedback, but negative
feedback is missing. This problem is also defined as the one-class problem.
There are multiple previous works aiming to deal with the one-class problem. A
simple solution is to treat all missing data as negative feedback. However, this is
not true because the missing (unobserved) interaction could be positive if the user is
aware of the item existing. Therefore, using this strategy to build a model might result
in a misleading model due to faulty assumptions at the outset. On the contrary, if
8 M. Alfarhood and J. Cheng

we treat all missing data as unobserved data without considering including negative
feedback in the model training, the corresponding trained model is probably useless
since it is only trained on positive data. As a result, sampling negative feedback
from positive feedback is one practical solution for this problem, which has been
proposed by [12]. In addition, Weighted Regularized Matrix Factorization (WRMF)
[9] is another proposed solution that introduces a confidence variable that works as
a weight to measure how likely a user is to like an item.
In general, the recommendation problem with implicit data is usually formulated
as follows:
1, if there is user-item interaction
Rnm = (8)
0, otherwise

where the ones in implicit feedback represent all the positive feedback. However,
it is important to note that a value of 0 does not imply always negative feedback.
It may be that users are not aware of the existence of those items. In addition,
the user-item interaction matrix (R) is usually highly imbalanced, such that the
number of the observed interactions is much less than the number of the unobserved
interactions. In other words, matrix R is very sparse, meaning that users only interact
explicitly or implicitly with a very small number of items compared to the total
number of items in this matrix. Sparsity is one frequent problem in RSs, which brings
a real challenge for any proposed model to have the capability to provide effective
personalized recommendations under this situation. The following sections explain
our methodology, where we aim to eliminate the influence of the aforementioned
problems.

3.2 The Attentive Autoencoder

Autoencoder [13] is an unsupervised learning neural network that is useful for com-
pressing high-dimensional input data into a lower dimensional representation while
preserving the abstract nature of the data. The autoencoder network is generally
composed of two main components, i.e., the encoder and the decoder. The encoder
takes the input and encodes it through multiple hidden layers and then generates a
compressed representative vector, Z j . The encoding function can be formulated as
Z j = f (X j ). Subsequently, the decoder can be used then to reconstruct and estimate
the original input, Xˆ j , using the representative vector, Z j . The decoder function can
be formulated as Xˆ j = f (Z j ). Each the encoder and the decoder usually consist of
the same number of hidden layers and neurons. The output of each hidden layer is
computed as follows:

h () = σ (h (−1) W () + b() ) (9)

Deep Learning-Based Recommender Systems 9

where () is the layer number, W is the weights matrix, b is the bias vector, and σ
is a non-linear activation function. We use the Rectified Linear Unit (ReLU) as the
activation function.
Our model takes input from the article’s textual data, X j = {x 1 , x 2 , . . . , x s },
where x i is a value between [0, 1] and s represents the vocabulary size of the arti-
cles’ titles and abstracts. In other words, the input of our autoencoder network is
a normalized bag-of-words histograms of filtered vocabularies of the articles’ titles
and abstracts.
Batch Normalization (BN) [14] has been proven to be a proper solution for the
internal covariant shift problem, where the layer’s input distribution in deep neural
networks changes across the time of training, and causes difficulty in training the
model. In addition, BN can work as a regularization procedure like Dropout [15]
in deep neural networks. Accordingly, we apply a batch normalization layer after
each hidden layer in our autoencoder to obtain a stable distribution from each layer’s
output.
Furthermore, we use the idea of the attention mechanism to work between the
encoder and the decoder, such that only the relevant parts of the encoder output are
selected for the input reconstruction. Attention in deep learning can be described
simply as a vector of weights to show the importance of the input elements. Thus,
the intuition behind attention is that not all parts of the input are equally significant,
i.e., only few parts are significant for the model. We first calculate the scores as the
probability distribution of the encoder’s output using the so f tmax(.) function.

ezc
f (z c ) = zd
(10)
de

The probability distribution and the encoder output are then multiplied using
element-wise multiplication function to get Z j .
We use the attentive autoencoder to pretrain the items’ contextual information
and then integrate the compressed representation, Z j , in computing the items’ latent
factors, V j , from the matrix factorization method. The dimension space of Z j and V j
are set to be equal to each other. Finally, we adopt the binary cross-entropy (Eq. 11)
as the loss function we want to minimize in our attentive autoencoder model.

L=− yk log( pk ) − (1 − yk ) log(1 − pk ) (11)
k

where yk corresponds to the correct labels and pk corresponds to the predicted

values.
10 M. Alfarhood and J. Cheng

3.3 Probabilistic Matrix Factorization

Probabilistic Matrix Factorization (PMF) [3] is a probabilistic linear model where

the prior distributions of the latent factors and users’ preferences are drawn from
Gaussian normal distribution.

u i ∼ N (0, λ−1
u I)
v j ∼ N (0, λ−1
v I) (12)
pi j ∼ N (u i v Tj , σ 2 )

We integrate the items’ contents, trained through the attentive autoencoder, into
PMF. Therefore, the objective function in Eq. 3 has been changed slightly to become
 ci j λu λv  
L= ( pi j − u i v Tj )2 + u i 2 + v j − θ (X j )2 (13)
i, j∈R
2 2 2

where θ (X j ) = Encoder (X j ) = Z j .
Thus, taking the partial derivative of our previous objective function with respect
to both u i and v j results in the following equations that minimize our objective
function the most

u i = (V Ci V T + λu I )−1 V Ci Pi
(14)
v j = (U C j U T + λv I )−1 U C j P j + λv θ (X j )

We optimize the values of u i and v j using the Alternating Least Squares (ALS)
optimization method.

3.4 Prediction

After our model has been trained and the latent factors of users and articles, U and
V , are identified, we calculate our model’s prediction scores of useri and each article
as the dot product of vector u i with all vectors in V as scor esi = u i V T . Then, we
sort all articles based on our model predication scores in descending order, and then
recommend the top-K articles for that useri . We go through all users in U in our
evaluation and report the average performance among all users. The overall process
of our approach is illustrated in Algorithm 1.
Deep Learning-Based Recommender Systems 11

Algorithm 1: CATA algorithm

1 pretrain autoencoder with input X ;
2 Z ← θ(X );
3 U, V ← Initialize with random values;
4 while <NOT converge> do
5 for <each user i > do
6 u i ← update using Equation 14;
7 end for
8 for <each article j > do
9 vi ← update using Equation 14;
10 end for
11 end while
12 for <each user i > do
13 scor esi ← u i V T ;
14 sort(scor esi ) in descending order;
15 end for
16 Evaluate the top-K recommendations;

4 Experiments

In this section, we conduct extensive experiments aiming to answer the following

research questions:
• RQ1: How does our proposed model, CATA, perform against state-of-the-art meth-
ods?
• RQ2: Does adding the attention mechanism actually improve our model perfor-
mance?
• RQ3: How could different values of the regularization parameters (λu and λv )
affect CATA performance?
• RQ4: What is the impact of different dimension values of users and items’ latent
space on CATA performance?
• RQ5: How many training epochs are sufficient for pretraining our autoencoder?
Before answering these research questions, we first describe the datasets used in our
evaluations, the evaluation metrics, and the baseline approaches we use to evaluate
our model against.

4.1 Datasets

Three scientific article datasets are used to evaluate our model against the state-of-
the-art methods. All datasets are collected from CiteULike website. The first dataset
is called Citeulike-a, which is collected by [5]. It has 5,551 users, 16,980 articles, and
204,986 user-article pairs. The sparseness of this dataset is extremely high, where
only around 0.22% of the user-article matrix has interactions. Each user has at least
12 M. Alfarhood and J. Cheng

ten articles in his or her library. On average, each user has 37 articles in his or her
library and each article has been added to 12 users’ libraries. The second dataset is
called Citeulike-t, which is collected by [6]. It has 7,947 users, 25,975 articles, and
134,860 user-article pairs. This dataset is actually sparser than the first one with only
0.07% available user-article interactions. Each user has at least three articles in his
or her library. On average, each user has 17 articles in his or her library and each
article has been added to five users’ libraries. Lastly, Citeulike-2004–2007 is the third
dataset, and it is collected by [16]. It is three times bigger than the previous ones with
regard to the user-article matrix. It has 3,039 users, 210,137 articles, and 284,960
user-article pairs. This dataset is the sparsest in this experiment, with a sparsity equal
to 99.95%. Each user has at least ten articles in his or her library. On average, each
user has 94 articles in his or her library and each article has been added only to one
user library. Brief statistics of the datasets are shown in Table 1.
Title and abstract of each article are given in each dataset. The average number
of words per article in both title and abstract after our text preprocessing is 67 words
in Citeulike-a, 19 words in Citeulike-t, and 55 words in Citeulike-2004–2007. We
follow the same preprocessing techniques as the state-of-the-art models in [5, 7,
8]. A five-stage procedure to preprocess the textual content is displayed in Fig. 3.
Each article title and abstract are combined together and then are preprocessed such
that stop words are removed. After that, top-N distinct words based on the TF-IDF
measurement are picked out. 8,000 distinct words are selected for the Citeulike-a
dataset, 20,000 distinct words are selected for the Citeulike-t dataset, and 19,871
distinct words are selected for the Citeulike-2004–2007 dataset to form the bag-of-
words histogram, which are then normalized into values between 0 and 1 based on
the vocabularies’ occurrences.

Table 1 Descriptions of citeulike datasets

Dataset #Users #Articles #Pairs Sparsity (%)
Citeulike-a 5,551 16,980 204,986 99.78
Citeulike-t 7,947 25,975 134,860 99.93
Citeulike-2004–2007 3,039 210,137 284,960 99.95

Fig. 3 A five-stage procedure for preprocessing articles’ titles and abstracts

Deep Learning-Based Recommender Systems 13

Fig. 4 Ratio of articles that have been added to ≤N users’ libraries

Figure 4 shows the ratio of articles that have been added to five or fewer users’
libraries. For example, 15, 77, and 99% of the articles in Citeulike-a, Citeulike-t, and
Citeulike-2004–2007, respectively, are added to five or fewer users’ libraries. Also,
only 1% of the articles in Citeulike-a have been added only to one user library, while
the rest of the articles have been added to more than this number. On the contrary,
13, and 77% of the articles in Citeulike-t and Citeulike-2004–2007 have been added
only to one user library. This proves the sparseness of the data with regard to articles
as we go from one dataset to another.

4.2 Evaluation Methodology

We follow the state-of-the-art techniques [6–8] to generate our training and testing
sets. For each dataset, we create two versions of the dataset for sparse and dense
settings. In total, six dataset cases are used in our evaluation. To form the sparse
(P = 1) and the dense (P = 10) datasets, P items are randomly selected from each
user library to generate the training set while the remaining items from each user
library are used to generate the testing set. As a result, when P = 1, only 2.7, 5.9,
and 1.1% of the data entries are used to generate the training set in Citeulike-a,
Citeulike-t, and Citeulike-2004–2007, respectively. Similarly, 27.1, 39.6, and 10.7%
of the data entries are used to generate the training set when P = 10 as Fig. 5 shows.
14 M. Alfarhood and J. Cheng

(a) Citeulike-a (b) Citeulike-t (c) Citeulike-2004-2007

Fig. 5 The percentage of the data entries that forms the training and testing sets in all citeulike
datasets

We use recall and Discounted Cumulative Gain (DCG) as our evaluation metrics
to test how our model performs. Recall is usually used to evaluate recommender
systems with implicit feedback. However, precision is not favorable to use with
implicit feedback because the zero value in the user-article interaction matrix has
two meanings: either the user is not interested in the article, or the user is not aware
of the existence of this article. Therefore, using the precision metric only assumes
that for each zero value the user is not interested in the article, which is not the case.
Recall per user can be measured using the following formula:

Relevant Articles ∩ K Recommended Articles

recall@K = (15)
Relevant Articles
where K is set manually in the experiment and represents the top K articles of each
user. We set K = 10, 50, 100, 150, 200, 250, and 300 in our evaluations. The overall
recall can be calculated as the average recall among all users. If K equals the number
of articles in the dataset, recall will have a value of 1.
Recall, however, does not take into account the ranking of articles within the
top-K recommendations, as long as they are in the top-K list. However, DCG does.
DCG shows the capability of the recommendation engine to recommend articles at
the top of the ranking list. Articles in higher ranked K positions have more value than
others. The DCG among all users can be measured using the following equation:
|U |
1   rel(i)
K
DCG@K = (16)
|U | u=1 i=1 log2 (i + 1)

where |U | is the total number of users, i is the rank of the top-K articles recommended
by the model, and rel(i) is an indicator function that outputs 1 if the article at rank i
is a relevant article, and 0 otherwise.
Deep Learning-Based Recommender Systems 15

4.3 Baselines

We evaluate our approach against the following baselines described below:

• POP: Popular predictor is a non-personalized recommender system. It recom-
mends the most popular articles in the training set, such that all users get identical
recommendations. It is widely used as the baseline for personalized recommen-
dation models.
• CDL: Collaborative Deep Learning (CDL) [7] is a deep Bayesian model that
jointly models both user-item interaction data and items’ content via a Stacked
Denoising Autoencoder (SDAE) with a Probabilistic Matrix Factorization (PMF).
• CML+F: Collaborative Metric Learning (CML) [17] is a metric learning model
that pulls items liked by a user closer to that user. Recommendations are then
generated based on the K-Nearest Neighbor of each user. CML+F additionally
uses a neural network with two fully connected layers to train items’ features
(articles’ tags in this chapter) to update items’ location. CML+F has been shown
to have a better performance than CML.
• CVAE: Collaborative Variational Autoencoder (CVAE) [8] is a probabilistic
model that jointly models both user-item interaction data and items’ content
through a Variational Autoencoder (VAE) with a Probabilistic Matrix Factoriza-
tion (PMF). It can be considered as the baseline of our proposed approach since
CVAE and CATA share the same strategy.
For hyper-parameter settings, we set the confidence variables (i.e., a and b) to
a = 1, and b = 0.01. These are the same values used in CDL and CVAE as well.
Also, a four-layer network is used to construct our attentive autoencoder. The four-
layer network has the following shape “#Vocabularies-400-200-100-50-100-200-
400-#Vocabularies”.

4.4 Experimental Results

For each dataset, we repeat the data splitting four times with different random splits
of training and testing set, which has been previously described in the evaluation
methodology section. We use one split as a validation experiment to find the optimal
parameters of λu and λv for our model and the state-of-the-art models as well. We
search a grid of the following values {0.01, 0.1, 1, 10, 100} and the best values on
the validation experiment have been reported in Table 2. The other three splits are
used to report the average performance of our model against the baselines. In this
section, we address the research questions that have been previously defined in the
beginning of this section.
16 M. Alfarhood and J. Cheng

Table 2 Parameter settings for λu and λv based on the validation experiment

Approach Citeulike-a Citeulike-t Citeulike-2004–2007
Sparse Dense Sparse Dense Sparse Dense
λu λv λu λv λu λv λu λv λu λv λu λv
CDL 0.01 10 0.01 10 0.01 10 0.01 10 0.01 10 0.01 10
CVAE 0.1 10 1 10 0.1 10 0.1 10 0.1 10 0.1 10
CATA 10 0.1 10 0.1 10 0.1 10 0.1 10 0.1 10 0.1

4.4.1 RQ1

To evaluate how our model performs, we conduct quantitative and qualitative com-
parisons to answer this question. Figures 6, 7, 8, and 9 show the performance of the
top-K recommendations under the sparse and dense settings in terms of recall and
DCG. First, the sparse cases are very challenging for any proposed model since there
is less data for training. In the sparse setting where there is only one article in each
user’s library in the training set, our model, CATA, outperforms the baselines in all
datasets in terms of recall and DCG, as Figs. 6 and 7 show. More importantly, CATA
outperforms the baselines by a wide margin in the Citeulike-2004–2007 dataset,
where it is actually sparser and contains a huge number of articles. This validates the
robustness of our model against data sparsity.

(a) Citeulike-a (b) Citeulike-t (c) Citeulike-2004-2007

Fig. 6 Recall performance under the sparse setting, P = 1

(a) Citeulike-a (b) Citeulike-t (c) Citeulike-2004-2007

Fig. 7 DCG performance under the sparse setting, P = 1

Exploring the Variety of Random
Documents with Different Content
indeed we always did on such occasions, each to act as executor to
the other, and having cordially shaken hands, lest an opportunity of
so doing should not occur again, we fell fast asleep.
I had slept, perhaps, an hour and a half, when I was awoke by the
voice of the orderly sergeant, who came to inform us that the
movement of the army was countermanded. I will not say whether
the intelligence was received as acceptable or the reverse; indeed, I
question whether we ourselves knew at the moment whether we
were relieved by the reprieve or the contrary. One thing, however, is
certain, that we did not sleep the less soundly from knowing that at
least to-morrow was secured to us, to be passed in a state of vigour
and vitality, though perfectly aware that the peril of a battle must be
encountered before long; and hence that it was really a matter of
very little moment whether it should take place now or a few days
hence.
On mustering next morning upon the parade-ground, we learned
that our intended operations were impeded only by the very bad
state of the by-ways. Though the rain had ceased for some days
past, such was the quantity which had fallen that no artillery could,
as yet, move in any other direction than along the main road. The
continuance of dry weather for eight-and-forty hours would,
however, it was calculated, remove this obstacle to our advance; and
hence every man felt that he had but a couple of days to count
upon. By good fortune, these days continued clear and serene, and
the justice of our calculations was in due time evinced.
 CHAPTER VIII.
The 8th and 9th of November passed over without any event
occurring worthy of recital. On the former of these days, indeed, we
had the satisfaction to see a French gun-brig destroyed by one of
our light cruisers, a small schooner, off the harbour of St Jean de
Luz. She had lain there, as it appeared, for some time, and,
apprehensive of falling into our hands, had ventured on that day to
put to sea; but, being observed by a brig, and the schooner above
alluded to, she was immediately followed, and after an action of
nearly an hour's continuance, she blew up. Whether her crew had
abandoned her previous to the explosion I had no means of
ascertaining.
Meanwhile, among ourselves, and throughout the different divisions
contiguous to us, a silence like that of a calm before a storm
prevailed. Each man looked as if he knew that an attack was
impending, though few conjectures were hazarded touching the
precise moment of its occurrence. On the evening of the 9th,
however, all doubt was at length removed. We were assembled at
parade, or rather the parade was dismissed, but, the band
continuing to play, the officers were waiting in groups about the
colonel's tent, when an aide-de-camp riding up informed us that the
army was to advance upon the morrow. The battalion to which I
belonged was appointed to carry the village of Urogne—a place
containing perhaps a hundred houses and a church; and we were to
take post for the purpose an hour before daybreak on the highroad,
close to the advanced sentinels. Of the disposition of other corps we
knew nothing.
As soon as the aide-de-camp departed, we began, as people so
circumstanced generally do, to discuss the propriety of our general's
arrangements. On the present occasion we were more than usually
convinced of the sagacity and profound skill of the noble lord. Our
corps had been selected, in preference to many others, for a service,
perilous, it is true, but therefore honourable. This showed that he
knew at least on whom he could depend; and we, of course, were
determined to prove that his confidence had not been misplaced.
Having passed an hour or two in this manner, we departed each to
his own tent, in order to make the necessary preparations for the
morrow. These were soon completed. Our baggage was packed; our
horses and mules, which, for the sake of shelter, had been kept
during the last ten days at certain houses in the rear, were called in;
and provisions enough for one day's consumption put into a
haversack. With this and our cloaks we directed a Portuguese lad, a
servant of Grey, to follow the battalion upon a little pony, which we
kept chiefly for such uses; and finally, having renewed our
directions, the one to the other, respecting the conduct of the
survivor in case either of us should fall, we lay down.
It was quite dark when I arose. Our fires had all burned out; there
was no moon in the heavens, and the stars were in a great measure
obscured by clouds; but we took our places instinctively, and in
profound silence. On these occasions I have been always struck with
the great coolness of the women. You seldom hear an expression of
alarm escape them; indeed they become, probably from habit, and
from the example of others, to the full as indifferent to danger as
their husbands. I fear, too, that the sort of life which they lead, after
they have for any length of time followed an army in the field, sadly
unsexes them (if I may be permitted to coin such a word for their
benefit); at least I recollect but one instance in which any symptoms
of real sorrow were shown even by those whom the fate of a battle
had rendered widows. Sixty women only being permitted to
accompany a battalion, they are, of course, perfectly sure of getting
as many husbands as they may choose; and hence few widows of
soldiers continue in a state of widowhood for any unreasonable
length of time; so far, indeed, they are a highly-favoured class of
female society.
The column being formed, and the tents and baggage so disposed
as that, in case of a reverse, they might be carried to the rear
without confusion or delay, the word was given to move on. As our
route lay over ground extremely uneven, we moved for a while
slowly and with caution, till, having gained the highroad, we were
enabled to quicken our pace. We proceeded thus, perhaps, about a
mile, when the watch-fire of a German picket was seen. Then the
order to halt being passed quietly from rank to rank, we ordered
arms, and sat down upon the green banks by the roadside. Here we
were directed to remain till there should be light enough to make
surrounding objects clear, and a gun somewhere to the left of us
should give the signal of attack.
Men are very differently affected at different times, even though the
situations in which they may be cast bear a strong affinity to one
another. On the present occasion, for example, I perfectly recollect
that hardly any feeling of seriousness pervaded my own mind, nor, if
I might judge from appearances, the minds of those around me.
Much conversation, on the contrary, passed among us in whispers,
but it was all of as light a character as if the business in which we
were about to engage were mere amusement, and not that kind of
play in which men stake their lives. Anxiety and restlessness, indeed,
universally prevailed. We looked to the east, and watched the
gradual approach of dawn; but it was with that degree of interest
which sportsmen feel on the morning of the twelfth of August—or
rather, perhaps, like that of a child in a box at Covent Garden, when
it expects every moment to see the stage curtain lifted. We were
exceedingly anxious to begin the fray, but we were quite confident
of success.
In the meanwhile, such dispositions were made as the circumstances
of the case appeared to require. Three companies, consisting of
about a hundred and fifty men, were detached, under the command
of a field-officer, a little to the right and left of the road, for the
purpose of surprising, if possible, two of the enemy's pickets. The
remaining seven, forming into column as day broke, extended their
front so as to cover the whole breadth of the road, and made ready
to rush at once, in what is called "double quick," upon the village.
That it was strongly barricaded and filled with French infantry we
were aware; but, by making our first attack a rapid one, we
calculated on reaching the barricade before the enemy should have
time to do us much damage by their fire.
We stood, perhaps, half an hour after these dispositions were
effected before the signal was given, the dawn gradually brightening
over the whole face of the sky. Now we could observe that we had
diverged in some degree from the main road, and occupied with our
little column a lane hemmed in on both sides by high hedges;
presently we were able to remark that the lane again united itself
with the road about a hundred yards in front of us; then the church
and houses of the village began to show through the darkness like
rocks or mounds; by-and-by the stubble-fields immediately around
could be distinguished from green meadows; then the hedgerows
which separated one field from another became visible. And now the
signal-gun was fired. It was immediately answered by a couple of
nine-pounders, which were stationed in a field adjoining to the lane
where we stood; and the battle began.
The three detached companies did their best to surprise the French
pickets, but without success, the French troops being too watchful to
be easily taken. They drove them in, however, in gallant style; and
the little column, according to the preconcerted plan, pressed
forward. Meanwhile the houses and barricade of Urogne were
thronged with defenders, who saluted us, as we approached, with a
sharp discharge of musketry. The bullets whistled round us, but with
less effect than might have been expected. A few men and one
officer fell, the latter being shot through the heart. He uttered but a
single word—the name of his favourite comrade—and expired. On
our part we had no time for firing, but rushed on to the charge;
whilst the nine-pounders already alluded to cleared the barricade
with grape and canister. In two minutes we had reached its base; in
an instant more we were on the top of it; when the enemy, as if
panic-struck at the celerity of our movements, abandoned their
defences and fled. We followed them through the street of the place,
as far as its extremity; but having been previously directed to
proceed no farther, we halted there, and they escaped to the high
grounds beyond.
The position now attacked in front of St Jean de Luz was one of
which Lord Wellington himself has said that he never beheld
anything more formidable. It extended for about three miles along
the ridge of a rising ground, the ascent of which was for the most
part covered with thick wood and intersected by deep ditches. These
natural defences Marshal Soult had strengthened with redoubts,
open batteries, and breastworks; the completion of which was begun
prior to our passage of the Bidassoa, and finished during our
compulsory halt on the heights above Handaye. Towards our left,
indeed—that is, towards the right of the enemy, and in the direction
of the village which we had just carried—the works in question
presented so commanding an appearance that our gallant leader
deemed it unwise to attempt any serious impression there; and
hence, having possessed ourselves of Urogne, we were directed to
attempt nothing more than to keep it at all hazards, and to make,
from time to time, a demonstration of advancing. This was done in
order to deter Soult from detaching any of this corps to the
assistance of his left, which it was the object of Lord Wellington to
overwhelm, and which, after twelve hours of severe fighting, he
succeeded in turning.
As soon as we had cleared the place of its defenders, we set about
intrenching ourselves, in case any attempt should be made to retake
the village. For this purpose we tore up the barricade erected by the
French, consisting of casks filled with earth, manure, and rubbish,
and rolling them down to the other end of the town, soon threw up
a parapet for our defence. The enemy, meanwhile, began to collect a
dense mass of infantry upon the brow of the hill opposite, and
turning a battery of three pieces of cannon upon us, they swept the
street with round-shot. These whizzing along, caused the walls and
roofs of the houses to crumble; but neither the shot, nor the shells,
which from time to time burst about us, did any considerable
execution. By avoiding conspicuous places, we managed to keep
well out of reach; and hence the chief injury done by the cannonade
fell upon the proprietors of houses.
We found in the village a good store of brown bread and several
casks of brandy. The latter were instantly knocked on the head and
the spirits poured out into the street, as the best means of hindering
our men from getting drunk; but the former was divided amongst
us; and even the black bread issued to the French soldiers proved a
treat to us, who had tasted nothing except biscuits, and these none
of the most fresh, for the last three months. We were not, however,
allowed much time to regale ourselves.
It was now about eleven o'clock, and the enemy had as yet made no
attack upon us. We could perceive, indeed, from the glancing of
bayonets through the wood in front, that troops were there
mustering; and as the country was well adapted for skirmishing,
being a good deal intersected with ditches, hedges, and hollow
ways, it was deemed prudent to send out three or four companies to
watch their movements. Among the companies thus sent out was
that to which I belonged. We took a direction to the left of the
village, and, being noticed by the enemy's artillery, were saluted
with a shower of round-shot and shell. Just at this moment a tumbril
or ammunition-waggon coming up, a shell from a French mortar fell
upon it. It exploded, and two unfortunate artillery drivers who
chanced to be sitting upon it were hurled into the air. I looked at
them a moment after they fell. One was quite dead, and dreadfully
mangled; the other was as black as a coal, but he was alive, and
groaned heavily. He lifted his head as we passed, and wished us
success. What became of him afterwards I know not, but there
appeared little chance of his recovery.
Having gained a hollow road somewhat in advance of the village, we
found ourselves in connection with a line of skirmishers thrown out
by Colonel Halket from his corps of light Germans, and in some
degree sheltered from the cannonade. But our repose was not of
long continuance. The enemy having collected a large force of
tirailleurs, came on with loud shouts, and every show of
determination. His object seemed to be to catch us in the hollow
way, where, because of the height and steepness of the bank, we
should have been at his mercy. The word was therefore passed to
move out and meet him, whereupon we clambered up the face of
the acclivity and dashed forward.
It would be hard to conceive a more animating military spectacle
than met the eye that day, as it moved to the right and left, tracing
the British line. For the benefit of my more peaceable readers, I may
as well mention that troops sent out to skirmish advance or retire in
files; each file, or pair of men, keeping about five yards from the
files on both sides of them. On the present occasion our line of
skirmishers extended about a mile in both directions, all spread out
in a sort of irregular order, and all firing independently of one
another, as the opportunity of a good aim presented itself. On the
side of the French, on the contrary, all was apparent confusion. Yet
the French tirailleurs are by no means in disorder when they appear
so. They are admirable skirmishers; and they gave our people this
day a good deal of employment before they again betook
themselves to the heights. They did not, however, succeed, as I
suspect was their design, in drawing us so far from the village as to
expose us to the fire of their masked batteries; but having followed
them across a few fields only, we once more returned to our hollow
road.
It was evident, from the numerous solid bodies of troops which kept
their ground along the enemy's front, that the plan of Lord
Wellington had been successful; and that no force had been sent
from the right of Soult's army to the assistance of his left. The
continuous roar of musketry and cannon which was kept up in that
direction proved, at the same time, that a more serious struggle was
going on there than any to which we were exposed. It was no rapid
but intermitting rattle, like that which we and our opponents from
time to time produced; but an unceasing volley, as if men were able
to fire without loading, or took no time to load. At length Soult
appeared to have discovered that he had little to dread upon his
right. About three o'clock we could observe a heavy column
beginning its march to the left; and at the same instant, as if to
cover the movement, the enemy's skirmishers again advanced.
Again we met them, as we had done before, and again drove them
in; when, instead of falling back to the hollow way, we lay down
behind a hedge, midway between the village and the base of their
position. From this they made several attempts to dislodge us, but
without effect; and here we remained till the approach of darkness
put an end to the battle.
The sun had set about an hour when the troops in advance were
everywhere recalled, and I and my companions returned to the
village. Upon it we found that the enemy still kept up an occasional
fire of cannon; and hence that the houses, which were extremely
thin, furnished no sufficient shelter for the troops. It was accordingly
determined to lodge the corps that night in the church; at the door
of which, to our great satisfaction, my friend and I found that our
Portuguese servant was waiting for us. The sumpter-pony was soon
unladen; and provisions and grog being at the same time served out
to the men, the graver business of the day was succeeded by
universal jollity and mirth.
The spectacle which the interior of the church of Urogne presented
that night was one which the pious founder of the fabric probably
never contemplated. Along the two side aisles the arms of the
battalion were piled, the men themselves occupying the centre aisle.
In the pulpit were placed the big drum and other musical
instruments, a party of officers taking possession of a gallery erected
at the lower extremity of the building. For our own parts, Grey and
myself asserted a claim to the space round the altar, which in an
English church is generally railed in, but which in foreign churches is
distinguished from the rest of the chancel only by its elevation. Here
we spread out our cold salt beef, our brown bread, our cheese, and
our wine; and here we ate and drank in that state of excited feeling
which attends every man who has gone safely through the perils of
such a day.
Nor was the wild nature of the spectacle around us diminished by
the gloomy and wavering light which thirty or forty small rosin tapers
cast over it. Of these, two or three stood beside us upon the altar.
The rest were scattered about by ones and twos in different places,
leaving every interval in a sort of shade, which gave a wider scope
to the imagination than to the senses. The buzz of conversation, too,
the frequent laugh and joke, and, by-and-by, the song, as the grog
began to circulate, all these combined to produce a scene too
striking to be soon forgotten.
As time passed on, all these sounds became more and more faint.
The men, wearied with their day's work, dropped asleep one after
another; and I, having watched them for a while, stretched out like
so many corpses upon the paved floor of the church, wrapped my
cloak round me, and prepared to follow their example. I laid myself
at the foot of the altar; and though the marble was not more soft
than marble usually is, I slept as soundly upon it as if it had been a
bed of down.
 CHAPTER IX.
We had slept about four or five hours, and the short hours of the
morning were beginning to be lengthened, when our slumbers were
disturbed by the arrival of a messenger from the advanced pickets,
who came to inform us that the enemy were moving. As we had lain
down in our clothes, with all our accoutrements on, we were under
arms and in column in five seconds. It was not, however, deemed
necessary that any immediate advance on our part should be
attempted. We remained, on the contrary, quiet in the church; but,
standing in our ranks, we were perfectly ready to march to any
quarter where the sound of firing might bespeak our presence
necessary.
We had stood thus about half an hour, when a second messenger
from the outposts came in, from whom we learned that a blue-light
had been thrown up within the enemy's lines, and that their fires
were all freshly trimmed. "Is it so?" said some of our oldest
veterans; "then there will be no work for us to-day—they are
retreating:" and so, sure enough, it proved. As soon as dawn began
to appear, a patrol was sent forward, which returned immediately to
state that not a vestige of the French army was to be found. Their
outposts and sentries were withdrawn, their baggage was all gone,
and the whole of the right wing had disappeared. The truth was, as
soon became apparent, that Lord Wellington's plan of attack had
succeeded at every point. The enemy's left was turned. His
redoubts, after some hard fighting, were taken; and our people
getting into his rear, left to Marshal Soult no alternative except
retreat. On his right, as I have already explained, no serious
impression could be made. There his position defied us. But our feint
deceived him into the hope that we might knock our heads against a
wall; and he delayed too long sending to the quarter where real
danger threatened the reinforcements that were needed. They
arrived in time to see that their vantage-ground was lost; and both
right and left withdrew, quietly and in good order, as soon as
darkness set in.
The intelligence of the enemy's retreat was received with infinite
satisfaction. Not that there was the smallest disinclination on our
part to renew the battle; but the battle being won, there remained
for us the rapture of the chase, than which, in the operations of war,
there is nothing more exciting. Very properly, however, the men
received orders to strengthen themselves against contingencies by
eating, before they began their march. This they did in the little
square or place whither the battalion was moved out of the church,
and round their piled arms—their breakfast, like our own, consisting
of such scraps of bread and meat as remained from the supper of
the previous evening.
While others were thus employed, I went, with two or three officers,
to visit the spot where we had deposited such of our messmates as
fell in the battle of yesterday. It is not often that a soldier is so
fortunate—if indeed the circumstance be worth a wise man's regard
—as to be laid to his rest in consecrated ground. Our gallant
comrades enjoyed that privilege on the present occasion. The
soldiers had collected them from the various spots where they lay,
and brought them in, with a sort of pious respect, to the churchyard.
Here they dug a grave—one grave, it is true, for more than one
body; but what boots it?—and here they entombed them, carefully
tearing up the green sod, and carefully replacing it upon the hillock.
For my own part, I had little time to do more than wish rest to their
souls; for the corps was already in motion, and in five minutes we
were in the line of march.
It was dark when the movement began, consequently objects could
not be distinguished at any considerable distance; but the farther we
proceeded, the more strongly the day dawned upon us. Having
cleared the village, we came to a bridge thrown across a little brook,
for the possession of which a good deal of fighting had taken place
towards evening on the previous day. Here we found several French
soldiers lying dead, as well as one of our own men, who had
ventured too far in pursuit. A little way beyond the bridge, and to
the left of the road, stood a neat chateau of some size. This our
advanced party was ordered to search; and as I chanced to be in
command of the detachment, the office of conducting the search
devolved upon me.
I found the house furnished after the French fashion, and the
furniture in a state of perfect preservation; nor did I permit the
slightest injury to be done to it by my men. The only article, indeed,
which I was guilty of plundering, was a grammar of the Spanish
language, thus entitled—"Grammaire et Dictionnaire François et
Espagnol—Nouvellement Revû, Corrigé et Augmenté par Monsieur
de Maunory: Suivant l'Usage de la Cour d'Espagne." Upon one of the
boards is written—Appartient a Lassalle Briguette, Lassallee. The
book is still in my possession; and as our countries are now at
peace, I take this opportunity of informing Mr Briguette that I am
quite ready to restore to him his property, provided he will favour me
with his address.
The room from which I took the volume just alluded to was the
library, and by no means badly stored with books. There was not,
however, much time to decipher the title-pages; for, independently
of the necessity under which we lay of pushing forward as soon as
we had ascertained that none of the enemy were secreted here, my
attention was attracted by a mass of letters scattered over the floor.
The reader may judge of my surprise, when, on lifting one to
examine its contents, I found it to be in the handwriting of my own
father, and addressed to myself. It was of a later date, too, than any
communication which I had received from home; and beside it were
lying about twenty others, directed to different officers in the same
division with myself. This let me into a secret. The house in which I
now stood had been the official headquarters of Marshal Soult. A
courier, who was bringing letters from Lord Wellington's
headquarters, had been cut off by a patrol of the enemy's cavalry;
and hence all our epistles, including sundry billets-doux from fair
maidens at home, had been subjected to the scrutiny of the French
marshal and his staff.
Leaving other epistles to their fate, I put my own in my pocket, and,
stuffing my volume of plunder into my bosom, pushed on. About a
hundred yards in rear of the chateau we arrived at the first line of
works, consisting of a battery for two guns, with a deep trench in
front of it. It was flanked both on the right and left by farmhouses,
with a good deal of plantation, and a couple of garden walls, and
would have cost our people no inconsiderable loss had we been
foolhardy enough to attack it. It was erected just upon the
commencement of the rising ground. On passing it we found
ourselves at the ascent of a bare hill, about the altitude, perhaps, of
Shooter's Hill, and not dissimilar in general appearance, the summit
of which was covered by three redoubts, connected, the one with
the other, by two open batteries. As we gazed at these, we could not
but remark to ourselves how painful must have been the feelings of
the French general when compelled to abandon them; and we, of
course, paid the compliments which were his due to our own leader,
by whose judicious management the labours of months were
rendered profitless to such an adversary.
We had just cleared the intrenchments when a cry arose from the
rear, "Make way for the cavalry!" Our men, accordingly, inclined to
the right of the road, when the 12th and 16th light dragoons rode
past at a quick trot, sending out half a troop before them to feel
their way. The object of this movement, as we afterwards found,
was to hinder, if possible, the destruction of the bridge at St Jean de
Luz; but the attempt succeeded only in part, the enemy having
already set fire to their train.
"Push on, push on," was now the word. We accordingly quickened
our pace, and reached St Jean de Luz about nine o'clock; but we
were too late to secure a passage of the Nivelle, the bridge being
already in ruins. Our cavalry had reached it only in time to see the
mine exploded, which the French troops had dug in its centre arch;
and hence a halt became necessary till the chasm thus created
should be filled up. The effect was remarkably striking; the whole of
the first and fifth divisions, with the King's German legion, several
brigades of Portuguese, and two divisions of Spanish troops, came
pouring up, till the southern suburb of St Jean de Luz was filled with
armed men, to the number of perhaps twenty or thirty thousand.
It is probably needless for me to say that we found St Jean de Luz
for the most part abandoned by its inhabitants. Here and there,
indeed, a few faces were protruded from windows and balconies,
and feeble cries of Vivent les Anglois! and the waving of threescore
or so of handkerchiefs, greeted our progress. But the persons so
conducting themselves belonged exclusively to the lower orders, for
the gentry and municipality were gone. It is just to add, however,
that in the course of a very few days both the gentry and municipal
authorities returned, and that they, and indeed all the inhabitants of
the place, were not only protected from insult and wrong, but
encouraged to resume, as they did, their usual occupations.
While the column halted till the bridge should be so far repaired as
to permit the infantry to cross, I happened to stray a little from the
main street, and beheld, in a lane which ran parallel with the river, a
spectacle exceedingly shocking. I saw no fewer than fifty-three
donkeys standing with the sinews of the hinder legs cut through. On
inquiring from an inhabitant the cause of this, he told me that these
poor brutes, being overloaded with the baggage of the French army,
had knocked up; when the soldiers, rather than suffer them to fall
into our hands in a serviceable condition, hamstrung them all. Why
they were not merciful enough to shoot them, I know not; unless,
indeed, they were apprehensive of causing an alarm among us by
the report. But what their caution hindered we performed: the poor
creatures were all shot dead ere we advanced.
The town of St Jean de Luz covers about as much ground, and, I
should imagine, contains about as many inhabitants, as Carlisle or
Canterbury. It is divided into two parts by the river Nivelle, which
falls into the sea about a couple or three miles below, at a village, or
rather port, called Secoa. Like other French towns of its size, St Jean
de Luz is not remarkable for its air of neatness; but there are in it a
good market-place, two or three churches, and a theatre. The
Nivelle, where it flows through the city, may be about the width of
the Eden or the Isis. It is rendered passable, and the two quarters of
the city are connected, by a stone bridge of three arches—beside
which, the stream itself is fordable, both for cavalry and infantry, at
low water. When we came in this morning the tide was up, but it had
been for some time on the turn; and hence, in about a couple of
hours, we were perfectly independent of the repairs. By this time,
however, the broken arch had been united by means of planks and
beams of wood; but as the junction was none of the most firm, it
was deemed prudent to send the troopers through the water, the
infantry only crossing by the bridge. Along with the cavalry was sent
the artillery also; and thus, by noon on the 11th of November, the
whole of the left column had passed the Nivelle.
We had hardly quitted St Jean de Luz when the weather, which
during the entire morning had looked suspicious, broke, and a cold
heavy rain began to fall. This lasted without any intermission till
dark, by which means our march became the reverse of agreeable;
and we felt as if we would have given the enemy a safe-conduct as
far as Bayonne, in return for permission to halt and dry ourselves
before a fire. But of halting no hint was dropped; nor was it till our
advanced-guard came up with the rear of the French army, posted in
the village of Bidart, and along the heights adjoining, that any check
was given to our progress. As it was now late, the sun having set
and twilight coming on, it was not judged expedient to dislodge the
enemy till morning; in consequence of which our troops were
commanded to halt. But there was no cover for them. Only a few
cottages stood near the road, and the tents were at least fourteen
miles in the rear; the night was accordingly spent by most of us on
the wet ground.
From the moment that the rain began to fall we remarked that the
Spanish, and in some instances the Portuguese troops, setting the
commands of their officers at defiance, left their ranks, and
scattered themselves over the face of the country. While this was
going on, I have good reason to believe that several horrible crimes
were perpetrated. Of the French peasants, many, trusting to our
proclamations, remained quietly in their houses; these were, in too
many instances, plundered and cruelly treated by the marauders,
who were, I suspect, urged on to the commission of numerous
atrocities by a feeling far more powerful than the desire of plunder. A
strong and overwhelming thirst of vengeance drew, I am convinced,
many to the perpetration of terrible deeds; indeed, one case of the
kind came under my own immediate notice, which I shall here
relate.
About three o'clock in the afternoon, a temporary check took place
in the line of march, when the corps to which I belonged was about
two miles distant from Bidart. A brigade of cavalry alone was in front
of us. A Portuguese brigade, including one regiment of caçadores,
was in our rear. We stood still in our places; but the caçadore
regiment, breaking its ranks, rushed in a tumultuous manner
towards two or three cottages on the left of the road. The officers,
with the utmost difficulty, recalled them; but a few individuals, as
the event proved, succeeded in their effort of insubordination.
These, however, were not noticed at the time, and it was thought
that all were where they ought to be.
A little way, perhaps a couple of hundred yards, in front, stood
another French cottage, surrounded by a garden, and detached from
all others. In about five minutes after order had been restored, we
heard a female shriek come from that cottage. It was followed by
the report of a musket; and ere we had time to reach the spot,
another shot was fired. We ran up and found a poor old French
peasant lying dead at the bottom of the garden. A bullet had passed
through his head, and his thin grey hairs were dyed with his own
blood. We hastened towards the house; and just as we neared the
door, a caçadore rushed out, and attempted to elude us. But he was
hotly pursued and taken. When he was brought back, we entered
the cottage, and, to our horror, we saw an old woman, in all
probability the wife of the aged peasant, lying dead in the kitchen.
The desperate Portuguese did not deny having perpetrated these
murders. He seemed, on the contrary, wound up to a pitch of frenzy.
"They murdered my father, they cut my mother's throat, and they
ravished my sister," said he; "and I vowed at the time that I would
put to death the first French family that fell into my hands. You may
hang me, if you will, but I have kept my oath, and I don't care for
dying." It is unnecessary to add that the man was hanged; indeed,
not fewer than eighteen Spanish and Portuguese soldiers were
tucked up, in the course of this and the following days, to the
branches of trees. But I could not at the time avoid thinking that if
any shadow of excuse for murder can be framed, the unfortunate
Portuguese who butchered this French family deserves the benefit of
it.
I have said that the greater part of the left column spent this night
in no very comfortable plight, upon the wet ground. For ourselves,
we were moved into what had once been a grass field, just at the
base of the hill of Bidart, but which, with the tread of men's feet and
horses' hoofs, was now battered into mud. Here, with the utmost
difficulty, we succeeded in lighting fires, round which we crowded as
we best might. But the rain still came down in torrents; and though
our lad arrived shortly after with the cloaks, and rations of beef and
biscuit and rum were served out to us, I cannot enumerate this
among the nights of pure enjoyment which my life as a soldier has
frequently brought in my way.
 CHAPTER X.
When I awoke next morning, I found myself lying in a perfect puddle
beside the decaying embers of a fire. The rain had come down so
incessantly and with such violence during the night, that my cloak,
though excellent of its kind, stood not out against it, and I was now
as thoroughly saturated with water as if I had been dragged through
the Nivelle. Of course my sensations were not of a very pleasant
nature; but I considered that I was far from singular in my condition,
and, like my comrades, I laughed at an evil for which there was no
remedy.
Having remained under arms till day had fully dawned, we began to
make ready for a further advance. When we lay down on the
preceding evening, several brigades of French troops were in
possession of Bidart. These we naturally laid our account with
attacking; but on sending forward a patrol, it was found that the
village had been abandoned, and that Soult had fallen back to his
intrenched camp in front of Bayonne. Our parade was accordingly
dismissed; and we remained in the same situation for about four
hours, when the arrival of the tents and baggage invited us to make
ourselves somewhat more comfortable. For this purpose the brigade
was moved about a quarter of a mile to the left of the main road;
and there, on a skirt of turf comparatively sound and unbroken, the
camp was pitched.
In the immediate vicinity of the tents stood a small farmhouse, or
rather a large cottage, containing three rooms and a kitchen. Thither
a good many of the officers, and myself among the number,
removed their canteens and portmanteaus, till no fewer than forty-
five individuals, including servants as well as masters, found a
temporary shelter under its roof. I am sure, after all, that I was not
more comfortable there than I should have been in my tent; but I
fancied that to sleep upon a bed once more, even though it was a
French one, would prove a luxury; and I made the experiment. It is
needless to add that the bed contained whole hordes of living
occupants besides myself, and that I did not venture again to
dispute with them the possession of their ancient domain.
From the 12th to the 17th November nothing occurred to myself, nor
were any movements made by the left of the British army worthy of
being related. The rain continued, with hardly any intermission,
during the whole of that time, rendering the cross-roads utterly
impassable for artillery, and holding out no prospect of fresh battles
or fresh adventures. It was indeed manifest that the troops could
not be kept much longer in the field without material injury to their
health, which began already to be threatened with dysentery and
ague. Nor is it surprising that the case should be so; for the tents
were not proof against showers so heavy and incessant as fell, and
canvas, when once completely soaked, allows water to pass through
it like a sieve. The consequence was that our men were never dry,
and many began to exhibit symptoms of the complaints above
alluded to.
Under these circumstances, we received with sincere rejoicing an
order in the evening of the 17th to strike our tents at dawn next
morning, and to march into winter quarters. The rain descended,
however, in such torrents, that though a temporary inconvenience
promised to lead to permanent comfort, it was deemed prudent to
delay fulfilling that order for at least some hours. We accordingly
remained quiet till about one o'clock in the afternoon of the 18th,
when, the weather breaking up, and the sun shining out, our camp
was struck, and we turned our faces towards the cantonments which
had been allotted to us.
Having cleared the few fields which intervened between the situation
of the camp and the highroad, we left Bidart behind, and took a
retrograde direction towards St Jean de Luz. We had not, however,
proceeded above five or six miles, and were still a full league distant
from the town, when, on reaching a cross-road which ran in a
direction to the left, we filed off and made for a piece of elevated
country, over which some half-dozen farmhouses were scattered.
These were assigned to the corps to which I belonged. We
accordingly halted on a sort of common near the centre of them,
and having cast lots as to which house should fall to the share of the
different companies, Grey, myself, and two others, with about a
hundred men, took possession of one, with which we were perfectly
satisfied.
It would be difficult for an ordinary reader to form any adequate
notion of the extreme satisfaction which soldiers experience when
first they establish themselves in winter quarters. As long as the
weather continues fine, and summer suns shed their influence over
it, there are indeed few places more agreeable than a camp. But it is
not so after the summer has departed. I have already hinted that
against heavy and continued rains a tent supplies but an inadequate
shelter. A tent is, moreover, but a narrow chamber, in which it is not
easy so much as to stand upright, excepting in one spot, and where
all opportunity of locomotion is denied. It furnishes, moreover, little
protection against cold—to light a fire within being impossible on
account of the smoke; and hence the only means of keeping yourself
warm is to wrap your cloak or a blanket about you, and to lie down.
Occasionally, indeed, I have seen red-hot shot employed as heaters;
but the kind of warmth which arises from heated iron is, at least to
me, hardly more agreeable than that which is produced by charcoal.
In a word, however enthusiastic a man may be in his profession, he
begins, about the end of October or the beginning of November, to
grow heartily tired of campaigning, and looks forward to a few
weeks' rest, and a substantial protection against cold and damps,
with almost as much pleasure as he experiences when the return of
spring calls him once more into the field.
The farmhouses in the south of France, like those in Spain, are
rarely provided with fireplaces in any other apartment than the
kitchen. It is, indeed, customary for families to live during the winter
months entirely with their servants; and hence the want of a
fireplace in the parlour is not felt any more than in the bedrooms. I
observed, likewise, that hardly any maison of the kind was furnished
with glazed windows, wooden lattices being almost universally
substituted. These, during the summer months, are kept open all
day, and closed only at night; and I believe that the extreme
mildness of the climate renders an open window at such seasons
very agreeable. On the present occasion, however, we anticipated no
slight annoyance from the absence of these two essential matters, a
chimney and a window in our room; and we immediately set our
wits to work to remove the evil.
Both Grey's servant and my own chanced to be exceedingly
ingenious fellows; the former, in particular, could, to use a vulgar
phrase, turn his hand to anything. Under his directions we set a
party of men to work, and, knocking a hole through one corner of
our room, we speedily converted it into a fireplace. To give vent to
the smoke, we took the trouble to build an external chimney,
carrying it up as high as the roof of the house; and our pride and
satisfaction were neither of them trifling when we found that it drew
to admiration. I cannot say that the masonry was very exact, or that
the sort of buttress added to the mansion improved its general
appearance; but it had the effect of rendering our apartment
exceedingly comfortable, and that was the sole object which we had
in view.
Having thus provided for our warmth, the next thing to be done was
to manufacture such a window as might supply us with light, and at
the same time resist the weather. For this purpose we lifted a couple
of lattices from their hinges, and having cut out four panels in each,
we covered the spaces with white paper soaked in oil. The light thus
admitted was not, indeed, very brilliant, but it was sufficient for all
our purposes; and we found, when the storm again returned, that
our oil-paper stood out against it stoutly. Then, having swept our
floor, unpacked and arranged the contents of our canteen, and
provided good dry hay-sacks for our couches, we felt as if the whole
world could have supplied no better or more desirable habitation.
To build the chimney and construct the window furnished occupation
enough for one day; the next was spent in cutting wood, and laying
in a store of fuel against the winter. In effecting this, it must be
confessed that we were not over-fastidious as to the source from
which it was derived; and hence a greater number of fruit-trees
were felled and cut to pieces than perhaps there was any positive
necessity to destroy. But it is impossible to guard against every little
excess when troops have established themselves in an enemy's
country; and the French have just cause of thankfulness that so little
comparative devastation marked the progress of our armies. Their
own, it is well known, were not remarkable for their orderly conduct
in such countries as they overran.
I have dwelt upon these little circumstances longer perhaps than
their insignificance in the eyes of my reader may warrant, but I
could not help it. There is no period of my life on which I look back
with more unmixed pleasure than that which saw me for the first
time set down in winter quarters. And hence every trifling incident
connected with it, however unimportant to others, appears the
reverse of unimportant to me. And such, I believe, is universally the
case when a man undertakes to be his own biographer. Things and
occurrences which, to the world at large, seem wholly undeserving
of record, his own feelings prompt him to detail with unusual
minuteness, even though he may be conscious all the while that he
is entering upon details which his readers will scarcely take the
trouble to follow.
Having thus rendered our quarters as snug as they were capable of
being made, my friend and myself proceeded daily into the adjoining
woods in search of game; and as the frost set in we found them
amply stored, not only with hares and rabbits, but with woodcocks,
snipes, and other birds of passage. We were not, however, so
fortunate as to fall in with any of the wild boars which are said to
frequent these thickets, though we devoted more than one morning
to the search; but we managed to supply our own table, and the
tables of several of our comrades, with a very agreeable addition to
the lean beef which was issued out to us. Nor were other luxuries
wanting. The peasantry, having recovered their confidence, returned
in great numbers to their homes, and seldom failed to call at our
mansion once or twice a-week with wine, fresh bread, cider, and
bottled beer; by the help of which we continued to fare well as long
as our fast-diminishing stock of money lasted. I say fast-diminishing
stock of money, for as yet no addition had been made to that which
each of us brought with him from England; and though the pay of
the army was now six months in arrear, but faint hopes were
entertained of any immediate donative.
It was not, however, among regimental and other inferior officers
alone that this period of military inaction was esteemed and acted
upon as one of enjoyment. Lord Wellington's fox-hounds were
unkennelled; and he himself took the field regularly twice a-week, as
if he had been a denizen of Leicestershire, or any other sporting
county in England. I need not add that few packs in any county
could be better attended. Not that the horses of all the huntsmen
were of the best breed, or of the gayest appearance; but what was
wanting in individual splendour, was made up by the number of
Nimrods; nor would it be easy to discover a field more fruitful in
laughable occurrences, which no man more heartily enjoyed than
the gallant Marquess himself. When the hounds were out, he was no
longer the Commander of the Forces, the General-in-Chief of three
nations, and the representative of three sovereigns, but the gay
merry country gentleman, who rode at everything, and laughed as
loud when he fell himself as when he witnessed the fall of a brother
sportsman.
Thus passed about twenty days, during the greater number of which
the sky was clear and the air cold and bracing. Occasionally, indeed,
we varied our sporting life by visits to St Jean de Luz, and other
towns in the rear, and by seeking out old friends in other divisions of
the army. Nor were we altogether without military occupation. Here
and there a redoubt was thrown up for the purpose of rendering our
position doubly secure; and the various brigades of each division
relieved one another in taking the duty of the outposts. A trifling
skirmish or two likewise tended to keep us alive; but these were
followed by no movement of importance, nor were they very fatal
either to the enemy or ourselves.
The position which Lord Wellington thus took up extended from the
village of Bidart on the left to a place called Garret's House on the
right. It embraced various other villages, such as that of Arcanques,
Gauthory, &c. &c., between these points, and kept the extremities of
the line at a distance of perhaps six or seven miles from each other.
To a common observer it certainly had in it nothing imposing, or
calculated to give the idea of great natural strength. On the left, in
particular, our troops, when called into the field, occupied a level
plain, wooded indeed, but very little broken; whilst at different
points in the centre there were passes easy of approach, and far
from defensible in any extraordinary degree. But its strength was
well tried, as I shall take occasion shortly to relate; and the issue of
the trial proved that no error had been committed in its selection.
Of the manner in which the right and centre columns were disposed
I knew but little. The left column, consisting of the first and fifth
divisions, of two or three brigades of Portuguese infantry, one
brigade of light and one of heavy cavalry, was thus posted: The
town of St Jean de Luz, in which Lord Wellington had fixed his
quarters, was occupied by three or four battalions of Guards; its
suburbs were given up to such corps of the German legion as were
attached to the first division. In and about the town the light cavalry
was likewise quartered; whilst the heavy was sent back to Handaye
and the villages near it, on account of the facility of procuring forage
which there existed. The Spaniards had again fallen back as far as
Irun, and were not brought up during the remainder of the winter;
but the Portuguese regiments were scattered, like ourselves, among
a number of detached cottages near the road. In the village of
Bidart was posted the fifth division, a battery or two of field-artillery,
and the men and horses attached to them; and to it the duty of
watching the enemy, and keeping possession of the ground on which
the pickets stood, was committed. Thus, along the line of the
highroad was housed a corps of about fifteen thousand infantry,
twelve hundred cavalry, and a due proportion of artillery—all under
the immediate command of Sir John Hope.
In direct communication with the head of this column was the light
division, under the command of Major-General Baron Alten. It
consisted of the 43d, 52d, and 95th regiments, of a brigade or two
of caçadores, and mustered in all about four or five thousand
bayonets. These occupied the church and village of Arcanques,
situated upon a rising ground, and of considerable natural strength.
Beyond this division again lay the fourth; in connection with which
were the third, the seventh, and the second divisions; whilst the
sixth took post a little in the rear, and acted as a reserve, in case a
reserve should be wanting.
I have said that Lord Wellington's headquarters were in St Jean de
Luz. Here also Sir John Hope, and several generals of division and of
brigade, established themselves; and here all the general staff of the
army was posted. Of course the place was kept in a state of warlike
gaiety, such as it had not probably witnessed before, at least in
modern times; but everything was done which could be thought of
to conciliate the goodwill of the inhabitants: nor was the slightest
outrage or riot permitted. Such is the manner in which the British
army was disposed of from the 18th of November, when it first went
into cantonments, till the 9th of December, when it was found
necessary once more to take the field.
 CHAPTER XI.
I had been out with my gun during the whole of the 8th of
December, and returned at a late hour in the evening, not a little
weary with wandering, when the first intelligence communicated to
me was that the corps had received orders to be under arms at an
early hour next morning, when the whole of the army would
advance. In a former chapter I have hinted that a continued tract of
rainy weather drove Lord Wellington, earlier than he had intended,
and against his inclination, into winter quarters. The consequence
was, that the position of the army was not in every respect to his
mind. The right, in particular, was too far thrown back; and the
course of the Nivelle interfered in a very inconvenient degree with
the communications between it and the left. We were accordingly
given to understand that the object of our present movement was
merely to facilitate the crossing of that river by Sir Rowland Hill's
corps; and that as soon as this object was attained, we should be
permitted to return in peace to our comfortable quarters.
In consequence of this information, Grey and myself made fewer
preparations than we had been in the habit of doing on other and
similar occasions. Instead of packing up our baggage, and ordering
out our sumpter-pony and faithful Portuguese, we left everything in
our apartment in its ordinary condition. Strict charges were indeed
given to the servants that a cheerful fire and a substantial meal
should be prepared against our return in the evening; but we put up
neither food nor clothes for immediate use, in full expectation that
such things would not be required.
The night of the 8th passed quietly over, and I arose about two
hours before dawn on the 9th, perfectly fresh, and, like those
around me, in high spirits. We had been so long idle, that the near
prospect of a little fighting, instead of creating gloomy sensations,
was viewed with sincere delight; and we took our places, and began
our march towards the highroad, in silence, it is true, but with
extreme goodwill. There we remained stationary till the day broke;
when the word being given to advance, we pushed forward in the
direction of Bayonne.
The brigade to which I belonged took post at the head of the first
division, and immediately in rear of the fifth. The situation afforded
to me on several occasions, as the inequalities of the road placed me
from time to time on the summit of an eminence, very favourable
opportunities of beholding the whole of the warlike mass which was
moving; nor is it easy to imagine a more imposing or more elevating
spectacle. The entire left wing of the army advanced in a single
continuous column along the main road, and covered, at the most
moderate computation, a space of four miles. As far, indeed, as the
eye could reach, nothing was to be seen except swarms of infantry,
clothed not only in scarlet, but in green, blue, and brown uniforms.
Here and there a brigade of guns occupied a vacant space between
the last files of one division and the first of another, and in rear of all
came the cavalry. Of their appearance I was unable accurately to
judge, they were so distant.
We had proceeded about five miles, and it was past eight o'clock,
when, our advanced-guard falling in with the French pickets, a smart
skirmish began. It was really a beautiful sight. The enemy made, it is
true, no very determined stand; but they did not give up a rood of
ground without exchanging a few shots with their assailants, who
pressed forward, vigorously indeed, but with all the caution and
circumspection which mark the advance of a skilful skirmisher. The
column, in the meanwhile, moved slowly but steadily on; nor was it
once called upon, during the whole of the day, to deploy into line.
When the light troops of an army are engaged as ours were this
morning, the heavy infantry marches at a slow rate; and short halts
or checks are constantly occurring. These befel to-day with unusual
frequency. The fact, I believe, was, that Lord Wellington had no
desire to bring his left into determined action at all. His object was
fully attained as long as he kept the right of the enemy in a state of
anxiety and irresolution; but the ground which we gained was in no
degree important to the furtherance of the sole design which we had
in view. Of course the tardiness of our motions gave a better
opportunity of watching the progress of those connected with us;
nor have I ever beheld a field-day at home more regularly gone
through than this trifling affair of the 9th December.
It was getting somewhat late—perhaps it might be three or four
o'clock in the afternoon—when our column, having overcome all
opposition, halted on some rising ground about three miles from the
walls of Bayonne. From this point we obtained a perfect view of the
outworks of that town, as well as of the formidable line of
fortifications which Soult had thrown up along the course of the
Adour; but of the city itself we saw little, on account of several
groves of elm and other trees which intervened. It will readily be
imagined that we turned our glasses towards the intrenched camp
with feelings very different from those which actuate an ordinary
observer of the face of a strange country. That the French marshal
had been at work upon these lines not only from the moment of his
last defeat, but from the very first day of his assuming the command
of the army of Spain, we were aware; and hence we were by no
means surprised at beholding such an obstacle presented to our
further progress in France. But I cannot say that the sight cast even
a damp upon our usual confidence. We knew that whatever could be
done to render these mighty preparations useless our gallant general
would effect; and perhaps we were each of us vain enough to
believe that nothing could resist our own individual valour. Be that as
it may, though we freely acknowledged that many a brave fellow
must find a grave ere these works could come into our possession,
we would have advanced to the attack at the instant, not only
without reluctance, but with the most perfect assurance of success.
The sound of firing had gradually subsided—the enemy having
withdrawn within their intrenchments, and our skirmishers being
called in to join their respective corps. The left column, dividing itself
according to its brigades, had taken post along a ridge of high
ground, and our men, piling their arms, set about lighting fires in all
directions, when I wandered from the corps, as my invariable
custom was, in search of adventures. I had strolled forward for the
purpose of obtaining, if possible, a more perfect view of the enemy's
lines, and was stepping across a ditch on my return, when a low
groan, as if from some person in acute pain, attracted my notice. I
looked down into the ditch, which was, perhaps, four feet deep, and
beheld three human beings lying at the bottom of it. They were all
perfectly naked, and two of them were motionless. On farther
examination, I found that they were three French soldiers, of whom
one only was alive; and he lay bleeding from a severe wound in the
face, a musket-ball having broken both cheek-bones. He was,
however, sensible; so I ran for help, and he was carried by some of
our people to a neighbouring house. Here the poor fellow, whom his
own countrymen had stripped and deserted, was well taken care of;
but he had suffered so much from exposure to cold that all attempts
to preserve life were vain, and he died in about a quarter of an hour
after his wound was dressed.
In the meanwhile Lord Wellington, putting himself at the head of a
small body of cavalry, and attended by a few companies of light
infantry, proceeded to the front in order to reconnoitre the enemy's
works. This he was permitted to do without any further molestation
than arose from the occasional discharge of a field-gun, as he and
his party presented a favourable mark to the gunners. But neither he
nor his followers received the slightest injury from these discharges;
and by six in the evening he had effected every object which he
desired to accomplish. Orders were accordingly issued for the troops
to fall back to their former quarters; and the main road was again
crowded with armed men marching to the rear, in a fashion not
perhaps quite so orderly as that which distinguished their advance.
A heavy rain had begun about an hour previous to this movement,
accompanied by a cold wind which blew directly in our faces;
darkness, too, set rapidly in; the road soon became deep and
muddy, from the trampling of a multitude of men and horses; and
something like an inclination to grumble began to arise in our
Welcome to our website – the ideal destination for book lovers and
knowledge seekers. With a mission to inspire endlessly, we offer a
vast collection of books, ranging from classic literary works to
specialized publications, self-development books, and children's
literature. Each book is a new journey of discovery, expanding
knowledge and enriching the soul of the reade

Our website is not just a platform for buying books, but a bridge
connecting readers to the timeless values of culture and wisdom. With
an elegant, user-friendly interface and an intelligent search system,
we are committed to providing a quick and convenient shopping
experience. Additionally, our special promotions and home delivery
services ensure that you save time and fully enjoy the joy of reading.

Let us accompany you on the journey of exploring knowledge and

personal growth!

ebooknice.com