Deep Learning Approaches towards Book Covers Classification
Przemyslaw Buczkowski
1,2
, Antoni Sobkowicz
2
and Marek Kozlowski
1
1
National Information Processing Institute, Warsaw, Poland
2
Warsaw University of Technology, Warsaw, Poland
Keywords:
Image Classification, Convolutional Neural Networks, Pattern Recognition, Deep Learning, Machine Lear-
ning, Supervised Learning, Artificial Intelligence.
Abstract:
Machine learning methods allow computers to use data in less and less structured form. Such data formats
were available only to humans until now. This in turn gives opportunities to automate new areas. Such
systems can be used for supporting administration of big e-commerce platforms e.g. searching for products
with inadequate descriptions. In this paper, we continue to try to extract information about books, but we
changed the domain of our predictions. Now we try to make guesses about a book based on an actual cover
image instead of short textual description. We compare how much information about the book can be extracted
from those sources and describe in detail our model and methodology. Promising results were achieved.
1 INTRODUCTION
In this paper, we focus on a task called image classi-
fication. The task is to predict the probability of clas-
ses given raw pixels intensities. Class refers to one
of predefined, discrete category. In machine learning
problems concerning image data, two distinct approa-
ches are available. First one, more classical, is to flat-
ten 2d image data into single dimension vector. Af-
ter such transformation, any classifier feedable with
the constant-length vector can be used. The problem
with this approach is that it ignores pixel neighbor-
hood and therefore cannot exploit information about
local patterns. In other words, this approach cannot
distinguish between pixels which are close to each ot-
her and those which are not, because proximity infor-
mation has been lost in the process (or in this repre-
sentation, to be more precise). Another, but related
flaw is a complete lack of invariance with respect to
translation because translating object few pixels away
creates big and hard-to-cope-with changes in the flat-
tened vector. On the other side are methods which
do not discard proximity information, quite the con-
trary, they are designed to exploit them. Firstly, it was
noticed that extracting spatial-aware features from an
image before using classifier may improve quality of
the model. Examples of such features are handcrafted
operators like Prewitt or Sobel (Adlakha et al., 2016).
A more advanced example is Gabor filter (Feichtinger
and Strohmer, 1998) which is inspired by reverse en-
gineering of visual cortex in the brain (Jones and Pal-
mer, 1987). Filters are small 2 dimensional images
which are multiplied pixel-wise with different parts
of the images. This operation is called convolution
and results with new, slightly smaller image. The
superior quality of systems based on convolution, or
more generally spacial–aware systems, caused an in-
crease of interest in such methods. Big image-related
machine learning challenges, like ILSVRC (Russa-
kovsky et al., 2015) or MSCOCO (Lin et al., 2014) are
dominated by convolutional neural network for a cou-
ple of years now (Krizhevsky et al., ; Szegedy et al., ;
He et al., ). In this paper, we successfully apply con-
volutional neural networks to book genre prediction
based on cover images. Dataset is acquired by craw-
ling from GoodReeds.com
1
. We briefly describe the
structure of this data set as well our previous analy-
sis with NLP methods. Lastly, we describe network
architecture and results.
2 RELATED WORK
Machine learning is found everywhere in today’s data
processing works. Most current machine learning
works well because of the human-designed represen-
tations and input features. Traditional methods are
mostly focused on numerical optimization of weig-
1
https://www.goodreads.com/
Buczkowski, P., Sobkowicz, A. and Kozlowski, M.
Deep Learning Approaches towards Book Covers Classification.
DOI: 10.5220/0006556103090316
In Proceedings of the 7th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2018), pages 309-316
ISBN: 978-989-758-276-9
Copyright © 2018 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
309

hts for human designed representations and featu-
res. Recently, representation learning, as sometimes
deep learning is called, has emerged as a new area
of Machine Learning research, and it attempts to au-
tomatically learn good latent features. Deep lear-
ning attempts to learn multiple levels of representa-
tion of increasing complexity/abstraction. The goal
of this approach is to explore how computers can take
advantage of data to develop features and represen-
tations appropriate for complex interpretation tasks.
The central idea behind early deep learning models
was to pre-train neural networks layer-per-layer in an
unsupervised fashion, which allows to learn hierar-
chy of features one level at a time. Moreover, pre-
training can be purely unsupervised, allowing rese-
archers to take advantage of the vast amount of un-
labeled data. Such approach makes Deep Learning
particularly well suited for Image and Natural Lan-
guage Processing tasks, where there is a huge number
of images and texts abound. Additionally, one can
use deep features as an input to standard supervised
machine learning methods.
Deep architectures are mainly neural networks
(recurrent, convolutional, deep belief) and can be
summarized as the composition of three elements: (1)
input layer - raw sensory inputs (e.g. words, Red-
Green-Blue values of pixels in an image); (2) hidden
layers - those layers learn more abstract non-obvious
representations/features; (3) output layer - predicting
the target (LeCun et al., 2015).
Recently, deep learning approaches have obtained
very high performance across many different NLP
tasks. These models can often be trained with a sin-
gle end-to-end model and do not require traditional,
task-specific feature engineering. The most attractive
quality of these techniques is that they can perform
well without any external hand-designed resources or
time-intensive feature engineering. Moreover, it has
been shown that unified architecture and learning al-
gorithm can be applied to solve several common NLP
tasks such as part-of-speech tagging, named entity re-
cognition or semantic role labeling (Collobert et al.,
2011). Such end-to-end system is capable of lear-
ning internal representation directly from the unla-
beled data allowing researchers to move away from
task-specific, hand-crafted features.
Similar insights are found in the image classifica-
tion and detection problems. In order to learn about
an enormous number of objects from millions of ima-
ges, we need a model with a large learning capacity.
However, the immense complexity of the object re-
cognition task means that this problem cannot be spe-
cified only by building such huge training data set,
so our model should also have lots of prior know-
ledge to compensate for all the data we dont have.
In particular, a deep convolutional neural network can
achieve reasonable performance on hard visual recog-
nition and categorization tasks – matching or excee-
ding human performance in some domains.
A Convolutional Neural Network (CNN) is a po-
werful machine learning technique from the field of
deep learning. CNNs are trained using large col-
lections of diverse images. Convolutional neural net-
work’s capacity can be controlled by varying their
depth and breadth, and they also make strong and
mostly correct assumptions about the nature of ima-
ges (namely, stationarity of statistics and locality of
pixel dependencies). Thus, compared to standard
feedforward neural networks with similarly-sized lay-
ers, CNNs have much fewer connections and para-
meters and so they are easier to train, while their
theoretically-best performance is likely to be only
slightly worse (Goodfellow et al., 2016). From these
large collections, CNNs can learn rich feature repre-
sentations for a wide range of images. These feature
representations often outperform hand-crafted featu-
res such as HOG, LBP, or SURF. An easy way to
leverage the power of CNNs, without investing time
and effort into training, is to use a pre-trained CNN as
a feature extractor for some multiclass linear SVM.
This approach to image category classification fol-
lows the standard practice of training an off-the-shelf
classifier using features extracted from images. For
example, the Image Category Classification Using
Bag Of Features example uses SURF features within a
bag of features framework to train a multiclass SVM.
The difference here is that instead of using image fe-
atures such as HOG or SURF, features are extrac-
ted using a CNN. Despite the attractive qualities of
CNNs, and despite the relative efficiency of their local
architecture, they have still been prohibitively expen-
sive to apply in large scale to high-resolution images
and usually, they demand GPUs grids to facilitate the
training of interestingly-large CNNs.
Researchers have demonstrated steady progress in
computer vision by validating their work against Ima-
geNet
2
– an academic benchmark for computer vi-
sion. Successive models continue to show impro-
vements, each time achieving a new state-of-the-art
result. ImageNet Large Visual Recognition Chal-
lenge is a standard task in computer vision, where
models try to classify entire images into 1000 clas-
ses, like ”Zebra”, ”Dalmatian”, and ”Dishwasher”. In
2012, an ensemble of CNNs achieved best results on
the ImageNet classification benchmark (Krizhevsky
et al., ). The authors of winning method trained a
large, deep convolutional neural network to classify
2
http://www.image-net.org/
ICPRAM 2018 - 7th International Conference on Pattern Recognition Applications and Methods
310

the millions of high-resolution images in the Image-
Net contest into the different classes. The neural net-
work, which has 60 million parameters and 650,000
neurons, consists of five convolutional layers, some
of which are followed by max-pooling layers, and
three fully-connected layers with a final 1000-way
softmax. To reduce overfitting in the fully-connected
layers recently-developed regularization method cal-
led dropout was used. They achieved a winning top-5
test error rate of 15.3%, compared to 26.2% achieved
by the second-best entry.
Also in 2012, the biggest NN so far (109 free pa-
rameters) was trained in unsupervised mode on un-
labeled images data (Le et al., 2012), then applied to
ImageNet. They trained a 9-layered locally connected
sparse autoencoder with pooling and local contrast
normalization on a large dataset of images (the model
has 1 billion connections, the dataset has 10 million
200x200 pixel images downloaded from the Internet).
The codes across its top layer were used to train a
simple supervised classifier, which achieved best re-
sults so far on 20,000 classes. Instead of relying on
efficient GPU programming, this was done by ”brute
force” on 1,000 standard machines with 16,000 cores.
So by 2011/2012, excellent results had been achie-
ved by Deep Learners in image recognition and classi-
fication. The computer vision community, however, is
especially interested in object detection in large ima-
ges, for applications such as image-based search en-
gines, or for biomedical diagnosis where the goal may
be to automatically detect tumors etc in images of hu-
man tissue. Object detection presents additional chal-
lenges.
3 PREVIOUS WORK
Our previous work is devoted to the issue of short
text classification, working on free textual descrip-
tions of books, gathered by crawling the GoodRe-
ads portal. Those descriptions are relatively short,
often incomplete and sometimes obscured by aut-
hor’s biographic note, which makes genre classifica-
tion a challenging task. There was a problem with
a huge amount of classes with most of them were
poorly and unevenly represented. We address these
issues more precisely in Data section as it was re-
levant to this paper as well. We compared two text
classification methods in order to choose the best
one for this specific task, including baseline naive
Bayes models and semantic enrichment method con-
suming neural-based distributional paragraph models
(Doc2Vec from gensim toolkit
3
). The sentence vector
3
https://radimrehurek.com/gensim/models/doc2vec.html
based methods – both original Doc2Vec (referred as
D2V) and averaged Doc2Vec category vectors (refer-
red as AD2V)– achieved much higher accuracy than
the baseline Multinomial Naive Bayes approach (re-
ferred as MNB) while requiring less text preproces-
sing. This difference points to the potential difference
in semantic context build around each book genre –
something that cannot be captured by a simple Baye-
sian classifier. The algorithms have been evaluated in
terms of the classification quality on the unique data
set of almost two hundred thousands book descripti-
ons. Results of previous work are presented in further
results section along with current results concerning
image-based categorization.
4 DATA
Figure 1: Examples of book cover images from our dataset.
Dataset consists of information crawled from website
GoodReads.com. GoodReads is a website which gat-
hers books-related information like reviews, recom-
mendations, and scores of over 160k books making
GoodReads basically the IMDB
4
for books. Most
data available is user-generated, including book genre
information. For most books textual as well as image
data is available. Information about the genre is avai-
lable as a list of pairs: genre name and number of user
votes. The first problem encountered is the huge num-
ber of categories which are poorly represented. There
are over 500 genres and most of them (over 400) are
extremely rare (less than 100 examples). This leads
to the need for limiting the number of genres so they
have better representation and balance. In order to do
so, we decided to assign every book a single ”main
category” by choosing genre with the most user votes
(we refer to this process later as ”relabelling”). Later
we sorted them by a number of those votes and picked
top 13 of them. All of the rest categories we labeled
as ”Other”. We ignored two categories ”Fiction” and
”Non-Fiction” as they can be considered as taxonomy
on a different (higher) level of hierarchy. According
to Wikipedia
5
Fiction and Non-Fiction consists of se-
parate genres.
4
http://www.imdb.com/
5
https://en.wikipedia.org/wiki/List of writing genres
Deep Learning Approaches towards Book Covers Classification
311

We are aware of some issues with the approach
we have chosen. Firstly we ignore nondominant vo-
tes for every book and assume a book has a single
category. This is convenient because it simplifies pro-
blem to simple classification. The gist of the problem
is that categories provided by users are not pairwise
disjoint as in classical classification problem. For ex-
ample, we can imagine a book which is Horror and
Romance at the same time. These categories, there-
fore, could be considered as tags and a single book
could have any number of predefined tags. Such pro-
blem could be addressed with many binary classifiers,
one classifier per tag. This would not be a perfect so-
lution as it only shifts problem somewhere else, na-
mely into settling threshold deciding how much votes
is required to assign a category. Another possibility
would be to discard some genres and keep only those
which cannot overlap. This is not a feasible task to
do because almost every real-world book is hybrid
of couple well-known genres. Even if it was possi-
ble it would be an arduous, manual and error-prone
process. Another problem is having genres which are
more subgenres than self-contained genres er genres,
e.g. Romance and Historical Romance. We didn’t
flatten those hierarchies mostly to be consistent with
our previous work in order to be able to compare re-
sults. Described issues (picking a single category and
keeping subcategories) are partially addressed by spe-
cial score function described later in the Evaluation
section.
Downloaded images have different sizes and ra-
tios. Most images have portrait ratio (bigger height
than width). Covers were scaled to fit 64x96 windows
and any empty spaces were filled with black pixels.
As this task requires pictures, records lacking ima-
ges were removed. This is the reason why number of
examples per categories changed since our first paper.
5 APPROACH
In this paper we focused on using convolutional neu-
ral network to predict book’s genre. We implemen-
ted relatively simple and shallow convolutional net-
work using TensorFlow
6
(Abadi et al., 2016). Ten-
sorFlow is Google’s open-sourced and unopinionated
framework for deeplearning. Our network consists
of three convolutional layers, each followed directly
by 2x2 max-pooling layer with non-overlapping win-
dows. Each convolutional layer has 3x3 kernels but
consists of increasingly more features maps: 16, 32
and 64. The stride of convolution in image space
6
https://www.tensorflow.org/
domain was set to 1, meaning convolution window
overlaps as much as possible. In all layers except
last layer Rectified Linear Unit (ReLU) activation
function (Glorot et al., 2011) was used as it yield su-
perior performance both during train and test phase.
ReLU given by (1)
ReLU(x) = max(0, x) (1)
is much simpler non-linear transformation than lo-
gistic function or hyperbolic tangent but it seems to
be common nowadays, especially with GPU compu-
tation. After those six layers, two fully connected lay-
ers are plugged in. First with 256 neurons and se-
cond is softmax (Sutton and Bart, 1998) layer with 14
neurons - typical for classification tasks. Softmax is
a function which amplifies maximal signal and dam-
pens others while normalizing outputs in such way
that they sum up to 1. Cross entropy error function
was minimized during training of the network (2).
crossEntropy(X ) = −
N
∑
i
L
∑
j
t
i j
log(p
i j
) (2)
Where t
i
j is equal to 1 if i − th example has class
j and 0 otherwise. p
i
j denotes probability (according
to model) that i − th example has category j. N and L
denotes the number of examples in dataset/batch and
number of classes/categories respectively. Stochas-
tic gradient descend optimization technique was app-
lied with RMSProp update rule(Tieleman and Hinton,
2012). This network will be referred to as N1.
Second, more sophisticated architecture was used
in network N2. The architecture of our second net-
work is inspired by VGG network (Simonyan and Zis-
serman, 2015). The main idea behind this kind of net-
work is to use several consecutive convolutional lay-
ers with small filters (3x3), put max-pooling layers af-
ter those convolutions and finish with an optional dro-
pout layer. Such block may be repeated multiple ti-
mes, usually with increasing number of feature maps.
We used blocks consisting of two consecutive con-
volutional layers with 3x3 filters, 2x2 max-pooling
layer with a stride of 2 and finally a dropout with pro-
bability parameter set to 0.25. Two of these blocks
were stacked sequentially: first with 32 feature maps
and second with 64. Dropout is a method focused on
preventing overfitting (Hinton et al., 2012; Srivastava,
2013). It is achieved by simply ignoring random part
of neurons during training of the model which leads
to more independent neurons without ”complex co-
adaptations”. This may be seen as a form of model
averaging as every batch is fed into slightly different
architecture. After second dropout layer, two dimen-
sional images are flattened into 1d vectors to fit first
ICPRAM 2018 - 7th International Conference on Pattern Recognition Applications and Methods
312

Figure 2: Genre distribution of 30 genres in the analyzed data set before relabelling. It is important to note that Fiction and
Non-fiction can be considered book types rather than genres.
of two fully connected layers. The first layer has 256
neurons and second 14. Between these two layers,
there is dropout layer with the probability set to 0.5.
Second of mentioned fully connected layer uses soft-
max activations similarly to N1. In this case, we used
SGD optimizer with Nesterov momentum (Nesterov,
2004) and small learning rate decay. Momentum met-
hods use information about previous displacements in
weight–space and apply them to current weight up-
date which prevents zig-zagging. Nesterov momen-
tum additionally uses special correction term to up-
date rule which provides better convergence. Decay
parameter decrease jump size as training progresses
which helps to approach minimum with better preci-
sion. Except for final softmax layer every layer use
ReLU activations.
During the training no data augmentation techni-
ques like rotating, flipping, cropping or elastic trans-
Deep Learning Approaches towards Book Covers Classification
313

formation were used. Due to data set imbalance bat-
ches were prepared in such way that every category is
represented by exactly the same number of examples.
This way network cannot get biased by mentioned im-
balance because it is not able to notice the difference
in a priori probabilities of genres.
6 EVALUATION
As previously stated in Data section we did not
choose to treat this problem as multi-label classifica-
tion as it is not obvious where to put threshold for
genres. This is the reason why we do not use any
of multi-label scores e.g. Hamming loss, Hamming
score or multi-label versions of precision and recall.
Instead we proposed custom score function which is
similar to TOP-k accuracy but uses weights to dis-
count rewards from not ideal prediction. In our so-
lution a book is assigned single dominant genre g
o
and classifier returns probabilities for each of 14 gen-
res g
ci
for i ∈ 1, 2, 3, ...14, where g
c1
denotes the most
probable genre, g
c2
second most probable genre and
so on. Reward from single book is given by s
g
(3).
s
g
=









1, if g
o
= g
c1
.
0.75, if g
o
= g
c2
.
0.5, if g
o
= g
c3
.
0.0, otherwise.
(3)
Our Score function is a simple average of all s
g
across dataset, where X denotes data set and N is a
size of that set.
Score(X) =
1
N
N
∑
i=1
s
g
(4)
7 RESULTS
Our smaller model N1 managed to achieve better
score (0.73) than more complex one N2 (0.68). The
accuracies of those models are 0.61 and 0.58 respecti-
vely. It is worth noticing that simpler network N1 not
only performs better but also converge faster in terms
of a number of epochs and time of training. Figure 3
shows the relation between score and epoch number.
We calculated our score function not only for
whole test set but also for particular categories. Those
scores are presented in Table 1. In that table, we also
included results of our previous models.
Table 1: Scores for evaluated classifiers: Multinomial Naive
Bayes (MNB), Doc2Vec (D2V) applied to textual descripti-
ons and two CNNs described in this paper N1 and N2 app-
lied to cover images.
Genre MNB D2V N1 N2
All genres 0.39 0.82 0.73 0.68
Mystery 0.28 0.84 0.27 0.29
Historical Romance 0 0.91 0.42 0.39
Young Adult 0.53 0.86 0.34 0.41
Science Fiction 0 0.83 0.27 0.24
Horror 0 0.87 0.17 0.10
Paranormal 0.01 0.91 0.19 0.11
Romance 0.71 0.91 0.37 0.45
Fantasy 0.93 0.87 0.44 0.74
Other 0 0.61 0.95 0.83
M M Romance 0 0.96 0.17 0.13
Historical Fiction 0.01 0.77 0.26 0.19
Classics 0 0.81 0.21 0.25
Manga 0 0.85 0.55 0.55
History 0 0.92 0.11 0.14
8 CONCLUSION
Networks achieved accuracy around 60% which can
be considered good in a 14-way classification pro-
blem. Considering properties of score function, score
around 0.75 means that on the average correct genre
is second most probable guess which is a pleasing re-
sult. This quality level qualifies proposed method to
be used in real-world system e.g. supporting workers
maintaining books catalog.
It is not surprising that predictions based on tex-
tual description are more accurate. There are some
keywords with big discriminative power. Such words
are likely to appear in some genre while having low
probability of occurrence in other genres. For ex-
ample ”frightening”, ”bloodcurdling” or ”chilling” in
horror books and ”charming” or ”lover” in romance
books. There is no such ”free dinner” in image dom-
ain. Covers are much more subtle. Some covers are
contradictory or not related to books’ content at the
first glance, requiring the reader to use complex so-
cial constructs or mentally solve some puzzle to un-
derstand a concept of the cover. Other covers might
be minimalist or mysterious to the level in which they
consist of single plain color and small title. In such
case human who does not know the language of the
title would be equally clueless as a convolutional neu-
ral network. We believe that considering relatively
small size of our dataset as well as complexity of
our visual models network can not spontaneously le-
arn features enabling it to recognize characters (or se-
quences of characters) and correlate them with genres
ICPRAM 2018 - 7th International Conference on Pattern Recognition Applications and Methods
314

Figure 3: Learning of networks. Left subplot depicts score of N1 network and the right subplot N2.
in a way NLP methods do.
We do not understand why neural networks find
some genres much more difficult than others. This
differences are much more obvious than in case of
NLP approaches. It looks like networks were trained
on different number of examples per category and le-
arned a priori distribution of genres. This can not be
the case as balancing method described earlier makes
sure every genre has the same number of examples
in each batch. Although every batch contains same
amount of each category, training set remained unba-
lanced (not trimmed) and therefore there were more
examples to sample from for some labels (e.g. other).
This could be the reason why decision boundaries in
decision space were set to favor ”other” as this label
would potentially cover greater area. Another expla-
nation for these differences in difficulty may be very
nature of cover images. For example intra-genre si-
milarity differs from genre to genre rendering recog-
nizing some of them much more difficult task. This
is even more plausible given that ”other” examples
consists of mixed, very specific genres (e.g. erotica,
vampires, biography).
9 FUTURE WORK
We plan to build ensemble classifier based on both
textual description and cover image and see if such
system built on top of these two existing systems will
yield substantially better results than any single sy-
stem. Right now we are building an online live demo
web application which we will share in the nearest fu-
ture. We also plan to try to use generative models to
synthesize textual description from cover images and
vice versa.
REFERENCES
Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z.,
Citro, C., Corrado, G. S., Davis, A., Dean, J., Devin,
M., et al. (2016). Tensorflow: Large-scale machine
learning on heterogeneous distributed systems.
Adlakha, D., Adlakha, D., and Tanwar, R. (2016). Analyti-
cal Comparison between Sobel and Prewitt Edge De-
tection Techniques. International Journal of Scientific
& Engineering Research.
Collobert, R., Weston, J., Karlen, L. B. M., Kavukcuoglu,
K., and Kuksa, P. (2011). Natural Language Proces-
sing (almost) from Scratch. Journal of Machine Lear-
ning ResearchVolume.
Feichtinger, H. G. and Strohmer, T. (1998). Gabor Analysis
and Algorithms.
Glorot, X., Bordes, A., and Bengio, Y. (2011). Deep sparse
rectifier neural networks.
Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep
learning. MIT Press.
He, K., Zhang, X., Ren, S., and Sun, J. Deep residual lear-
ning for image recognition. 2015.
Hinton, G., Srivastava, N., and Krizhevsky, A. (2012). Im-
proving neural networks by preventing co-adaptation
of feature detectors.
Jones, J. P. and Palmer, L. A. (1987). An evaluation of the
two-dimensional Gabor filter model of simple recep-
tive fields in cat striate cortex. Journal of Neurophy-
siology.
Krizhevsky, A., Sutskever, I., and Hinton, G. E. Image-
net classification with deep convolutional neural net-
works. 2012.
Le, Q. V., Ranzato, M., Monga, R., Devin, M., Corrado, G.,
Chen, K., Dean, J., and Ng, A. Y. (2012). Building
high-level features using large scale unsupervised le-
arning.
LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep lear-
ning. Nature, 521.
Deep Learning Approaches towards Book Covers Classification
315

Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P.,
Ramanan, D., Doll
´
ar, P., and Zitnick, C. L. (2014).
Microsoft coco: Common objects in context.
Nesterov, Y. (2004). Introductory Lectures on Convex Op-
timization. A Basic Course. Springer.
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S.,
Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bern-
stein, M., et al. (2015). Imagenet large scale visual
recognition challenge.
Simonyan, K. and Zisserman, A. (2015). Very deep convo-
lutional networks for large-scale image recognition.
Srivastava, N. (2013). Improving neural networks with dro-
pout.
Sutton, R. S. and Bart, A. G. (1998). Reinforcement Lear-
ning: An Introduction. MIT Press.
Szegedy, C., Liu, W., and Lia, Y. Going deeper with convo-
lutions. 2015.
Tieleman, T. and Hinton, G. (2012). Lecture 6.5 - rms-
prop, coursera: Neural networks for machine learning.
Technical report.
ICPRAM 2018 - 7th International Conference on Pattern Recognition Applications and Methods
316