Reinforcement Learning is the key to Artificial

General Intelligence
Ryan Haider

Student at Bella Vista High School, president of school Artificial Intelligence Club.

Abstract: Hello folks! I decided to write this paper to help you, reader/reviewer will start to know what you are really up to in your
someone who is like I was before, to finally learn how to write paper. The Introduction is, besides the Abstract, the
a machine learning paper. For this, I will report here all my
knowledge acquired in about 10 years of research, with the
best thesis and best journal paper awards on my back,
reviewing machine learning papers for several journals and
writing and researching for three different countries in three
different con- tinents. So let us start with the Abstract! the
abstract is the most important part of your scientific paper, as
it is the first text the readers will read about your work when
they search for it in digital libraries. In the abstract, you
summarize your work by talking about (i) the problem you
want to solve; (ii) why it is im- portant; (iii) how others deal
with the same problem; (iv) what are the limitations of other
works; (v) how do you solve such lim- itations and what is the
novelty of your work; and (vi) are the results promising? what
did you learn from the experiments? please be aware that the
abstract has a limitation on the num- ber of words. You need
to check it on the journal/conference website.
Here you put the keywords, which are words that best describe your work and
will also be used to find your work in digital libraries.
Correspondence: A paper has a corresponding author, who is the person
that will deal with all procedures related to the submission and publication
of the paper. Usually, the corresponding author is your advisor or the guy
who has the financial support to pay for the publication (if that’s the case),
or sometimes, this can also be the first author. Here you put the name and
e-mail address of the corresponding author.

1. Introduction. Hello, world! I wrote this

paper to de- fine a "standard" for the papers written
by my students and, this way, minimize the rejection
of papers that are considered badly written. So, I hope
it can help you too, especially if you are beginning
now in scientific life. My intention here is not to
define a silver bullet approach that can work all the
time and must be followed forever but make the
reviewers’ job of rejecting papers (sic) as difficult as
possible. So, with experi- ence, you will find your
paper format, but I am sure that most of what is
written here must remain in your paper. Please be
aware that such steps I am presenting here are only
for re- search papers that present a novel approach, so
Tutorials and Overview papers should not follow the
steps here. I chose a paper format that is a standard
for a given journal style, but the rules I describe here
can be applied to any journal or conference format.
Please be aware that journals and confer- ences have
limitations on the number of pages and number of
references, so you don’t want to have a paper rejected
for such a stupid reason, correct?

Tip: be careful with the Introduction!!!!. OK, now your

Anselmo Ferreira | UNISI-DIISM | September 27, 2022 | 1–8
most dangerous and important section of any
paper. One sim- ple English mistake here can cost
you a fast rejection. So, be a perfectionist in this
section!!!

What to write in the Introduction. Your

introduction is an extended version of the
Abstract, where you can discuss in further detail
the problem you want to solve and how you want
to solve it. So, be aware of the following
discussions that MUST be present in the
Introduction:
1. What is the problem you want to solve?
2. Why such a problem is important?
3. Is there any important impact for the
community if solving such a problem? do
the media discuss its im- portance? it would
be very nice to cite something from a news
media company about this problem to make
it closest to the reality of the
reader/reviewer.
4. how does the literature touch such a
problem? what are the solutions proposed
so far? what are their limi- tations in terms
of the proposed solution, datasets, and
difficulty of the experiments?
5. how is your proposed approach supposed to
defeat such limitations? what are you
proposing? what is the novelty? how does it
work? are the results promising? are there
any difficulties in the experiments you are
considering here but were not considered
previously?
After discussing these points, it is highly
recommendable to make it CLEAR, in a new
paragraph, the contributions of the paper in a
numbered list. For example, check the text below:
"In summary, the contributions of this paper are"
1. Contribution one
2. Contribution two
3. etc.
About this last issue, I like to use odd numbers (at
least three) to do such a contribution list. Maybe I
am superstitious about it (LOL).
So, after presenting the problem, confirming its
importance, telling the limitations of the related
works, and showing your novelties, it is time to
finish the Introduction by telling what is coming
next in the paper. You can start it in the following
way: "The remaining of this paper is organized as
follows. Section II discusses BLABLABLA,
Section III presents BLABLABLA, and so on."

2 | UNISI-DIISM Anselmo Ferreira | WritePaperDummies

Related Work in the Introduction?. This is something I and hold
don’t like to do, but it happens in some papers I have been
re- viewing in my life. Some people do a very exhaustive
related work discussion in the Introduction, which can make
the sec- tions of the paper unbalanced in terms of size (as a
reviewer, I don’t expect that the Introduction Section is
bigger than the Proposed Method section, for example). So,
my advice here is to cite the related work in a batch
depending on their sub- division (if the related work can be
subdivided into different branches), telling about their
limitations in general in a very brief manner. Then, in a
separate section, you discuss them in more detail. Anyway,
it is your option, but be aware of how big will be your
Introduction and how unbalanced your paper will be in the
end.

Identifying important features of a paper. Figures, ta-

bles, sections, and equations are written with the first letter
in capital or not, depending on different situations, as the
ex- amples below show:
• Figure/Table/Section/Equation 1 shows that...
• From this figure/table/section/equation we can see
that...

Use of numbers in a paper. This is probably a myth, but

one day somebody told me that, if I want to talk about a
num- ber that is less than 10, I must use letters to write
about it. Otherwise, I use numbers. Look at the examples
below.
• Five-fold cross-validation
• 10-fold cross-validation
Is that true? I am curious about it too. Check with your
advisor about it and tell me what you found!

2. Background and Related Work (optional). I advise

you to write a section like this. Sometimes, the reviewers
are Ph.D. students who are using the opportunity to review
to learn better the concepts of the area they want to act, or
sometimes they are forced by their supervisors to do it for
the journal they are editors (sic). Additionally, making your
work as auto-contained as possible will maintain the atten-
tion of the reader to your paper, without stopping reading it
to find basic concepts anywhere. So, your acceptance
chances for the reviewers and interest by the readers will be
higher. It is also good for you, because now YOU WILL
EXPLAIN THE PROBLEM with your own words, so you
will learn bet- ter by teaching (BTW, teaching something is
the best way of learning it, It’s proven by science;-)).
You can use both of them in the same section or divide
them into two different sections (background comes first). I
did both of these options for my papers.

Background. Here you will discuss the basic concepts of

the environment you are aiming to act. For example, in the
fi- nancial market, there are a series of procedures that the
Arti- ficial Intelligence (AI) developer must know before
applying an AI solution to it. For example, what are buy
Anselmo Ferreira | UNISI-DIISM | September 27, 2022 | 1–8
 later.

The Related Work. OK, so here we start talking about so-

lutions that acted in a problem similar to ours (or something
similar if you are solving a new problem never tackled be-
fore). Please, pay attention to the fact that I have witnessed

Fig. 1. This is a very nice picture a co-author of mine did to explain

how laser printers work. It is based on a Wikipedia article we saw
before. Please notice that this is not plagiarism, saying that is the
same as Apple trying to patent rectangles because Apple’s
smartphones are rectangular (sic). Don’t forget to explain, in a
very succinct way, what is the picture you are showing about.

operations? what are long and short operations?

how is the stocked market data available to the
general public? what is time series? so, different
environments contain different concepts that must
be explained to the reader, so she/he can understand
the environment you want to act. One very nice
thing to do is, depending on the environment you
want to talk about, put a very nice picture to explain
it.

About Pictures in a Scientific Article. Please, be

super aware of using somebody else’s pictures from
another paper in your paper. These pictures are
usually copyrighted by the publisher, so
SOMETIMES you need to ask permission from the
authors AND the publisher to use the figure in your
pa- per. Another solution is YOU BUILDING THE
PICTURE YOURSELF, based on the one you
wanted to use in a given image editor. Be aware that
this is not plagiarism, as a picture explaining an
environment can be drawn the same way by
different people. You just need to build it yourself
in a given image editor and, of course, do some
modifications to don’t make them the same picture.
Take a look at Figure 1, where my co-author based
his drawing on another one present in the literature.

What About Different Backgrounds?.

Sometimes, be- sides the background about the
application, you also have the background of the
mathematical model you are using as a solution for
the problem (e.g., you want to talk about basic
concepts of deep learning). I don’t like to put such
different backgrounds in the same section, but you
can do it. When it’s the case for me, I usually talk
about the basic concepts of the environment in the
Introduction and the basic concepts of the solution
in this section. Another option would be talk- ing
about the background of the environment in this
section and the background of the solution in the
Proposed Method Section, which shall be discussed
4 | UNISI-DIISM Anselmo Ferreira | WritePaperDummies
lots of papers rejected for a very stupid reason, which is an this)
outdated related work discussion. So, my algorithm to
elimi- nate such a problem is:

• try to discuss something like 20 related works;

• consider the two last years, plus the current year;

• balance your references in conference and journal pa-

pers;

• consider old references only if they are classical solu-

tions, with a super high number of citations; and

• use a diverse number of publishers with high reputa-

tion in your related works (e.g., IEEE, Elsevier,
ACM, Springer, among others).

Sometimes, in your review process, the reviewers will sug-

gest you consider SOME SPECIFIC ARTICLES EVEN IF
THEY ARE OLD OR LESS CITED. If that happens, BE
SURE THAT THE AUTHOR OF THE PAPER IS THE
RE-
VIEWER OF YOUR PAPER. So, do what they say and try
to understand their approach, explain it as best as possible,
and try to be kind with the limitations of their approach
(LOL). By the way, they work for free so they want their
papers cited and usually use the review process for that
(#SadBut- True). The same can happen when sometimes
the reviewers or editors ask you to include more references
from their jour- nal. Just do what they say and they will be
happy (#SadBut- True2.0).

Abbreviations in Machine Learning Papers. If you are

a machine learning enthusiast, you might hear about terms
such as SVMs (Support Vector Machines), and PCA
(Princi- pal Component Analysis) among others. However,
how do you know that the reader of your paper is an
enthusiast like you? so, I don’t like to see what I call
undeclared variables in machine learning papers. I know
that some terms become boring to repeat in the text, but
every time you think about using an abbreviation, don’t
forget to declare what that ab- breviation means before in
your text and use it as you want. In My Humble Opinion
(IMHO), this will grab the reader’s attention to your text,
and he/she will not need to look at what that ******
abbreviation means anywhere. So, IMHO, keep your
abbreviations declared in your text (you just need to do it
once before you start using them).

The Use of Latin words and Identifying a List of Au-

thors. Maybe you saw in your life the use of the Latin
word etc., but when you read related works papers you will
find other words, such as:

1. e.g., this is a Latin word acronym that means exempli

gratia, which translates to, literally, "for example."
Pe- riods come after each letter and a comma
normally fol- lows unless the example is a single
word and no pause is natural. I like to use it in
parenthesis to give exam- ples of something (e.g., like
Anselmo Ferreira | WritePaperDummies UNISI-DIISM | 3
 2. i.e., this means id est, which translation to are my tips:
English is "that is". Loosely, "i.e." is used to
mean "therefore" or "in other words." 1. The Introduction and Abstract of a paper are the
Periods come after each letter and a comma places most probable to help you understand
normally follows or not, depending on somebody’s method. So, read them and write in your
whether the wording following the own words what you understood from them. What
abbreviation dictates a natu- ral pause. I also does that ap- proach have that differentiates it from
like to use them in parenthesis (i.e., to say the others?
the same thing in other words).

3. et al. from the Latin et alii, which literally

means "and others". It must always be typed
with a space between the two words, and
with a period after the "l" (since the "al." is
an abbreviation). A comma does not follow
the abbreviation unless the sentence’s
grammar requires it.

Be aware that some journals italicize these terms

because they come from Latin, but most do not. I
usually put them in italics.
About the last word et al., it is used to identify a
list of authors with three or more authors. So, one
and two authors are usu- ally identified by their
last names in capital letters, and three and more
authors are identified by the first author’s surname
followed by the et al. word. Check the examples
below:

• RONALDO [1] proposed to celebrate his

goals by spinning in the air.

• SIMON and GARFUNKEL [2] composed a

very nice song about the sounds of silence.

• GIBB et al. [3], also known as the Bee

Gees, wrote something about staying alive.

Please, be aware that sometimes the authors’

names are not cited in papers, just their reference
numbers. Check what is the case with the journal
you want to submit.

How to search for related work. There are

several tools from different publishers to help you
find papers related to the application you are
working on. In the list that follows I will indicate
some websites that were active at the time I was
writing this paper:

1. Elsevier: https://www.sciencedirect.com/

2. IEEE: https://ieeexplore.ieee.org/Xplore/home.jsp

3. ACM: https://dl.acm.org/

4. Springer: https://link.springer.com/

You can use the same keywords as yours (do you

remember them?) in these tools to find similar
papers, or just put a phrase that can be easily
found on them. To describe the related work, here
4 | UNISI-DIISM Anselmo Ferreira | WritePaperDummies
 2. If you can divide the related works into branches, do to identify the source of a
it. For example, what are active and passive
approaches? what are the proposed approaches for
each of them?

3. Look for a limitation that you are exploring in your

paper but was not touched by the literature before. It
is easy to find, even your approach has limitations! so
what are the limitations of the literature solutions that
will be faced by your approach?

I like to finish the related works with that last issue. I can use
it as a link to the next section, where I explain my approach
and tell how I will deal with the literature limitations.

3. Proposed Method. If the reviewer got here without

find- ing any problem in your paper, congratulations! you
have now about 20% chance of having your paper accepted!
so, as I told you in the previous section, I like to finish the
related work with the limitations of the existing methods,
and I start the proposed method section by telling HOW I
AM TACK- LING EXISTING SOLUTIONS
LIMITATIONS. So, I start
this section by giving an overview of my approach, telling
how it works in a general way.
Ok, pay attention to this. There are two ways of grabbing the
reviewer’s attention to your paper:

1. what the reviewer READS.

2. what the reviewer SEES.

Can you realize that both things are done with the eyes?
which one do you think is the best? yes, me too, so, I will
teach you several tricks to grab the reviewer’s attention to
your text without using regular text. I call them attention
tools and they are the following:

1. Figures

2. Subsections

3. Equations

4. Algorithms

Let’s check them one by one in the next subsections.

Figures. You know that a picture is worth a thousand words,

correct? so, what if you could motivate and explain your
approach with beautiful pictures of the problem and your
so- lution? reviewers and readers will surely like it.
Although I understand that not all applications will allow us
to do that (e.g., it is very hard to understand the different
stock market behaviors at the same time and put in just one
graph), in some cases we can do that. Suppose that the
problem you have has a Gaussian distribution, so you can
plot such a Gaussian dis- tribution to show the reviewer that
your approach has a rea- son to be, and then you show a
picture of how your approach will deal with such a problem.
Nice, don’t you think? for example, in Figure 1, I showed
how a printer works. Now, I will show how a printer prints
characters to motivate my texture descriptor developed used
Anselmo Ferreira | WritePaperDummies UNISI-DIISM | 5
 from the problem you want to solve

• Approach (or pipeline) Pictures: will show how your

Fig. 2. Here is an example picture to motivate your approach. These are

micro- scope pictures of letters printed by different printers. It can be
seen from this figure that there are differences in textures between the
same letter printed by these dif- ferent printers. So, with this picture, I
motivate my solution without showing it to the reviewer.

printer. Check Figure 2, it’s a nice way to convince

the re- viewer that a printer can be identified before
telling about my printer source attribution approach,
don’t you think?
Ok, now you have done so far the following steps: (i)
iden- tified the limitations of existing approaches;
(ii) explained your approach in general terms; and
(iii) motivated your so- lution. So, it’s time to
finally present your approach to the reviewer. So,
how to do that? WITH ANOTHER FIGURE, OF
COURSE! I show you in Figure 3 a figure for
another application I wrote a paper, check from the
figure that it is possible to understand the approach
just by seeing that pic- ture.
OK, together with this figure you should also discuss,
in other words, how the approach works because, in
the case the fig- ure is not clear enough, you are
discussing it. You can also add more figures to
show details of your approach, or some piece of
your figure that requires more details. I usually do
this in subsections of my papers.
So, in summary, the types of figures you can use at the
begin- ning of your proposed method section are the
following:

• Motivation Pictures: to show which kind of

behavior have you noticed about the data
6 | UNISI-DIISM Anselmo Ferreira | WritePaperDummies
 IN SAMPLE DATASET OUT OF SAMPLE DATASET where λ is the frequency of the sinusoid factor, θ is the orien-
HYPERPARAMETERS OPTIMIZATION INTRINSIC PARAMETERS OPTIMIZATION
tation of the normal to the parallel stripes of the Gabor func-
Individual Classifiers
Intrinsic Parameters
Individual Classifiers
Intrinsic Parameters
tion, σ is the variance of the smooth curve (envelope) that
Train Val Test
Optimization
(local metric)
Train Val Test
Optimization
(local metric) outlines the extremes of the signal, and γ is the spatial as-
Classifier
1
Train Val Test Classifier
1
Train Val Test pect ratio which specifies the Gaussian ellipticity. Finally,
... ... Ensemble Hyperparameters
Optimization
... ... parameters λ and σ specify the resolution of the descriptor.
By varying the parameters λ and σ, the Gabor filters act
(global metric)

Train Val Test

con- sidering multiple resolutions.

OPTIMIZED
Ensemble Train Val Test INTRINSIC Final Result
PARAMETERS
...
Train Val Test You can notice two things about the equation above: (i) it is
Ensemble Train Val Test
part of the same paragraph that calls it; and (ii) it is finished
OPTIMIZED HYPERPARAMETERS
with a comma or period, which will indicate if the following
text is part of the same paragraph or not. Remember again:
Fig. 3. A pipeline of how to generate auto-configurable ensembles to time series data
classification. You can see that the approach has two steps, and different
all symbols must be defined.
parameters are optimized in different kinds of data. This kind of figure is interesting Although there is not a rule of thumb, I don’t expect a ma-
because it has several steps, so I can create subsections, algorithms, formulas, and chine learning paper with less than five equations (of
even more figures about each of them.
course, that can be alleviated for conference papers, but try
to follow this rule as much as you can).
approach works in general, with a given number of pro-
cedures dealing with the input and generating output. Algorithms. Here we arrived at the grand finale, or the
apotheosis of your proposed approach section, where you
Don’t forget to always explain these figures without specific will finally give as many details as possible of your
details at the beginning of your method’s section. approach. Here you will say how to transform your idea into
Now you can start to use the other attention tools to make code. Here, you will do what the scientific community in
your reviewer and reader interested in your approach. We machine learn- ing begs for researchers to do: MAKE
will do this in the following. YOUR WORK RE- PRODUCIBLE. I show you in
Algorithm 1 one example of doing that.
Subsections. This is an awesome trick. If the reviewer has
doubts about a piece of your pipeline picture he/she read at Algorithm 1 Proposed hyperparameter search approach
the beginning of the section, you are already making avail- Require:
able further details of it in an easy-to-find section of your 1: IS=time series from in sample data
2: I= list of intra-parameters
paper. I usually split pieces of my approach into subsec- 3: H=list of hyperparameters
tions to detail how they work. Another solution I do is us- 4: C=list of classifiers from the ensemble
Ensure:
ing one subsection to give basic concepts of the approach, 5: h'= Optimized hyperparameters
or the mathematical foundation behind it. I usually do it in 6: procedure RETURN_HYPERPARAMETERS(IS, I, H, C)
7: M AX_F INAL_M ETRIC ← 0
the Background section if the technique is very famous 8: ENS_METRIC ← 0
(e.g., deep learning) and use the Proposed Method Section 9: for h in H do . for each hyperparameter combination
10: W [h] ← buildW alks(IS, h(window_size)) .
to only show how I designed and dispose it in the proposed Starts non-anchored WF0
pipeline. Then, of course, I give details about the pipeline in 11: for w in W [h] do . for each walk
12: F ← buildF eatures(w, h(lags)) . get
other sub- sections. features 13: F ' ← icaT ransform(h(ica_comp), F ) . transform
features 14: M AX_W ALK_M ETRIC ← 0
15: for c in C do . for each classifier
16: for i in I do . for each intrinsic parameter, train and validate
Equations. Your approach, even if it’s not new, has some 17: '
M [i] ← trainClassifier(F , h(train_size), c[i])
mathematical foundation. So yes, you need to put equations 18: METRIC ←
testClassifier(M [i], F '[h(train_size) ∗ 0.3])
in your text. There are lots of possibilities to do that: do you 19: if METRIC > M AX_W ALK_M ETRIC then
transform your data? which calculations do you do after 20: E[c, w] ← M [i]
pre- processing? is there any normalization of the output? 21: M AX_W ALK_M ETRIC ← METRIC
22: end if
how does your classifier work? so, answering all these 23: end for
24: end for
questions will help you to find the right equations for your 25: test_data ← F'[h(window_size) − h(train_size) −
text. h(train_size) ∗ 0.3]
Please be aware that all the Equations of your paper must 26: ENS_METRIC ← ENS_METRIC +
testClassifier(E[C, w], test_data)
have their symbols defined after they are presented. Look at 27:
28:
end for
if ENS_METRIC > M AX_F IN AL_M ETRIC then
one example to do below: 5 3
1 (x cos θ + y 29: h' ← h
sin θ)2
30: M AX_F INAL_M ETRIC ← ENS_METRIC
ge(x, y) = exp − + (1) 31: end if
(−x 2 σ2 46
32: end for
sin θ + y cos 33: return h'
θ)2 × 34: end procedure
(γσ)2

Anselmo Ferreira | WritePaperDummies UNISI-DIISM | 7

 3 4
2π
cos x cos θ + y sin θ , The Algorithm you show here is just a confirmation of what
λ the reader or reviewer already knows about your approach
(2)

8 | UNISI-DIISM Anselmo Ferreira | WritePaperDummies

 from the previous attention tools, but now you discuss in to- Set of experiments #1: Preliminary Experiments. This
tally specific detail the approach. So, it is good to explain part is usually fused with the Proposed Method section, but
what the approach does line by line of the algorithm. To fin- I really like to do it in this section to make the reviewer
ish the proposed method section, a recommendable action to curious before and surprised later. The preliminary
do is tell the complexity of the algorithm, in order also to experiments are experiments related to your approach only
tell the reviewer about your approach in terms of running and are used to show the reviewer that what you are
time. Doing this, I believe this section is done. proposing works. Some kinds of preliminary experiments
are:
4. Experimental Setup. I always like to tell about the
infor- mation of the experiments in a specific section, even 1. Varying parameters of the proposed approach:
though I know that some people do this inside the sup- pose that one part of your approach deals with a
Experiments sec- tion. Although it’s not mandatory in differ- ent input than everything that was done before.
general, here are some topics I like to discuss in this section So, you experiment (you can use training and
validation data for that) to show the difference in
• Datasets built, considering the same level or different performance your approach gets with these different
levels of difficulty. inputs. I show in Table 1 one preliminary experiment
I’ve done that proves that given CNNs architectures
• Metrics used to assess the performance of the work better with pre-specified inputs for remote
proposed approach. sensing image classi- fication. You can also do here
an experiment to tell what happens with and without
• Methodology used for experiments (cross dataset, your proposed feature selection, feature
cross-validation, etc.) transformation, or pre-processed ap- proach. An
interesting thread that is been followed by writers
of Neural Network papers is to do what is called
• Experimental scenarios, telling the difficulty of each
ABLATION STUDY. In this study, "pieces" of the
experiment.
network are cut from the main network and their
effects in terms of the performance of the re-
• Other approaches (baselines) proposed before in the
maining network are evaluated.
literature that will compete with my proposed
approach in the experiments. I usually start declaring 2. Importance of features: you can use a Random
each com- petitor with an abbreviation (e.g., Principal Forest classifier to tell the importance of features, plot
Component Analysis, or PCA). DON’T it and discuss it. Check Figure 4 I did to identify the
IMPLEMENT THEM best features in my proposed feature set to detect
YOURSELF! Send a message to the authors of the blurred images.
machine learning paper you are interested in, and
ask for the code. I witnessed a case where an author 3. T-SNE plot: you can use the T-SNE approach to plot
implemented a baseline method via reverse engineer- your features and also the competitors in different
ing and got a paper rejected because, according to the sub- figures. Then, you can discuss how good your
reviewer’s opinion, the way he coded was wrong. ap- proach is to generating features that are easily
This is subjective, so avoid this kind of problem. sepa- rable by a machine learning classifier. I show in
Figure 5 how I did that to compare my approach with
• Statistical tests that will be done, to declare that my another.
approach is not winning by luck.
4. Hyper-plane Plot: this one I have seen in several ma-
chine learning tutorials and I am sure this would be a
• Implementation aspects of the approaches considered very nice thing to do in a paper (I have never done
for the experiments, including the programming lan- that before). If you show that your data is very
guage used, hardware, parameters of the classifiers, separable by the separating hyper-plane of a classifier
etc. like an SVM, the reviewer will certainly like it.

5. Experiments. Here we are, we will finally convince

the reviewer to accept our paper in this section. Here, you Metrics calculated considering PROPOSED_SUBMODEL1
Rank INPUT after classifying bag_2 images
will report the experiments done to prove that your
F NACC (%) TPR (%) FPR (%)
approach de- serves respect. However, I’ve been noticing 1 NEAR INFRARED CHANNEL 0.47 84.10 94.26 26.06
FALSE COLOR 0.34 74.52 92.90 43.86
that most of the papers just focus on showing that their 2
3 GREEN CHANNEL 0.29 67.00 81.91 47.90
approach simply per- forms their tasks better than the 4 RED CHANNEL 0.24 60.47 96.89 75.15
5 BLUE CHANNEL 0.20 50.02 100.00 99.96
others, and let me tell you, from my experience this is not
enough. There are a series of diverse experiments to be
Table 1. This is a preliminary experiment I did to show which is the best input for my
done to make the reviewer better convinced about the CNN approach to be applied in a specific kind of remote sensing image.
benefits of your approach. I will discuss them in the
following:
Anselmo Ferreira | WritePaperDummies UNISI-DIISM | 9
Main Experiments. NOW IT’S TIME TO SHINE!
here you will report the results considering your
approach and the

10 | UNISI- Anselmo Ferreira | WritePaperDummies

DIISM
 Statistics Calculated on CMEN Dataset after 5X2 Cross-Validation Experiments
Rank Method
F-MEASURE (%) ACC (%) TPR (%) FPR (%) PRECISION (%)
1 MULTISCALE BKS-RF- LVT 89.96 90.76 ± 14.36 82.95 ± 28.41 1.43 ± 3.15 98.31
2 BKS-RF-LVT 88.72 89.76 ± 14.51 81.10 ± 28.73 1.57 ± 2.85 98.10
3 MULTISCALE BKS-SVR-LVT 87.12 88.52 ± 15.67 78.03 ± 31.27 0.98 ± 2.27 100.00
4 BKS-SVR- OTSU 86.49 87.95 ± 13.06 77.35 ± 25.38 1.44 ± 3.34 98.17
5 BKS-SVR- LVT 83.64 85.91 ± 16.28 72.55 ± 32.55 0.72 ± 1.60 100.00
6 SURF [33] 80.40 83.50 ± 19.97 67.89 ± 39.41 0.87 ± 2.42 100.00
7 BKS [5] 79.52 82.76 ± 17.57 66.00 ± 35.18 0.48 ± 1.32 100.00
8 SIFT [33] 75.49 80.03 ± 21.30 60.63 ± 42.96 0.56 ± 1.19 100.00
9 Multiscale Voting [34] 74.56 78.33 ± 17.77 59.45 ± 37.24 2.79 ± 5.23 99.97
10 THRESHOLD VOTING (T=4) 68.15 75.73 ± 19.42 51.69 ± 38.82 0.22 ± 1.01 100.00
11 Zernike2 [48] 66.35 73.98 ± 20.74 49.65 ± 40.36 1.69 ± 3.30 99.98
12 Zernike [14] 62.38 72.56 ± 16.79 45.33 ± 33.39 0.20 ± 1.35 100.00
13 DCT [6] 54.22 68.53 ± 16.79 37.19 ± 33.55 0.13 ± 0.53 100.00
14 KPCA [13] 48.51 65.93 ± 15.66 32.02 ± 31.32 0.14 ± 0.77 100.00
15 THRESHOLD VOTING (T=6) 42.48 63.46 ± 17.02 26.97 ± 34.00 0.04 ± 0.43 100.00
16 Hierarch-SIFT [31] 39.51 61.99 ± 17.72 24.62 ± 35.70 0.64 ± 2.56 100.00
17 BAYESIAN FUSION [4] 8.48 52.20 ± 2.01 4.43 ± 4.02 0.03 ± 0.16 100.00

Legend:
xx.xx = Five best methods in the column metric
xx.xx = Five worst methods in the column metric

Table 2. Example of Table of results I did in Excel and converted to pdf, using it as an image in my paper. I usually consider abbreviations to call my approach and the
competitors and I leave my approaches in bold. I also highlight the best metrics in bold and use different colors to show the top-5 or bottom-5 metrics.

Scale 3 Scale 5 Scale 7 Scale 9

3 5 7 9

i i+1 i+2
Perturbations

Fig. 4. I use in this figure the Random Forests output that shows the importance of
each feature. As different parts of my proposed feature set are created by different
steps of the algorithm, I can say what steps are more important in my approach. Fig. 5. I used the T-SNE approach to show that my feature set (left) generates two classes of
features (red and blue) in the N-dimensional space that are in clusters far away from each
other. The same does not happen with the baseline approach (right). This facilitates the job of
binary classification techniques if applied to my proposed approach.

Anselmo Ferreira | WritePaperDummies UNISI-DIISM | 11

Fig. 6. This is an Equity Curve, or Profit and Loss curve done by my
co-author. This curve is used to compare how much money our
proposed machine approach earns through time in the stock market,
compared with a baseline approach (BH).

competitors from the literature in a real-world

application. You can use the following tools to do
that:
• Tables, to report metrics of your approach
and litera- ture solutions, ranked by a given
metric. I like to build tables in Excel and
convert them to pdf, using figures as tables.
However, as I like to put citations by the
side of the literature approaches
abbreviations on that table, sometimes I
need to change the table if the references
change, so, choose what is the best for you.
Look at my example in Table 2.
• Figures, such as the Receiver Operational
Curve and many others. Look at Figure 6 to
see one example.
In this part of the experiments, different
difficulties can be considered in different
subsections (for example, results of the approach
with and without attacks). It is a good practice to
also show samples that were misclassified by the
baselines, but classified perfectly by your
approach and explain why. I do this as an example
in Figure 7.

12 | UNISI- Anselmo Ferreira | WritePaperDummies

DIISM
 your paper will be rejected because precision is what
matters in most of the applications and as I said before,
what is slow today can be accelerated tomorrow according to
Moore’s law. Take a look at Table 4, where I do such a
running time com- parison that shows how slow is my
approach.

Mean Running Time Per Image

APPROACH (s)
Fig. 7. Some 64×64 patches representing regions from two different genuine 2D
SIFT [33] 2.43
barcodes misclassified by the baseline approach, but correctly classified by the pro- posed
approach. The difficulty of such a problem is highlighted by the fact that the halftones in
ZERNIKE [14] 42.08
the genuine class can assume multiple sizes, without any fixed pat- tern, which can be ZERNIKE2 [48] 2025.85
confused with irregular and additional edges from counterfeited barcodes if the descriptor BKS [5] 2893.75
used is not scale-invariant. MULTISCALE BKS-RF-LVT 2955.91
PROPOSED_SUBMODEL2
Table 4. Running time of my proposed approach (in bold) and competitors. Please
PROPOSED_FUSION

notice that my approach is composed of running eight approaches in a sequence

that could be run in parallel. So, such a result can be misleading.
GLCM-MD [60]
VGG-19 [16]
TMH3 [66]

TOTAL
Rank

Method 6. Conclusion. Wow! it’s finally done! we finished our

six-sections paper and now all we need to do is conclude
our work! although there is no rule of thumb for conclud-
ing something, here is a step-by-step approach I do in this
section:
1 PROPOSED_FUSION 0 1 1 1 1 4 1. I motivate the problem and discuss existing solutions
2 PROPOSED_SUBMODEL2 -1 0 1 1 1 2 limitations again, but in a very short paragraph.
3 TMH3 [66] -1 -1 0 1 1 0
4 GLCM-MD [60] -1 -1 -1 0 1 -2 2. I summarize what I am proposing in the paper and
5 VGG-19 [16] -1 -1 -1 -1 0 discuss what I discovered from experiments results,
-4
and difficulties, and even discuss drawbacks of my
1 = Line method is better than column method ap- proach. Additionally, I let it clear what are the
0 = Line method is equivalent to column method contri- butions my paper is giving to new researchers
-1 = Line method is worse than column method in the area
Table 3. A statistical test table that shows who wins, who loses, and if there is 3. I discuss future work that can be inspired by my re-
any statistical tie between each pair of approaches I considered for remote sensing
image classification.
search reported in the paper, such as modifying other
parameters, trying different classifiers, making the
Statistical Tests. In this kind of experiment, you use the datasets more difficult, and even studying the
chosen statistical test that you discussed in the Experimental possibil- ity of transferring my solution to another
Setup section, do you remember? then, you discuss here the problem.
results of such a test. I usually put a beautiful table here, to That’s all folks! my intention here is, again, to try to help
show if my approach wins the others in terms of confidence beginners to learn an initial format for their papers. I truly
level and sum up the scores, making a kind of leaderboard believe that other formats can be considered, but similar
in the end. If you decide to do this and are performing ideas can be found in this tutorial. Additionally, I believe
several experiments, I suggest you do this for all that other research areas can also use some of my concepts
experiments OR the most important and difficult one OR here to help them. Finally, the drawback of this tutorial is
for the experiments that your approach wins with small that it is not recommended for other paper formats, such as
metrics differences. Take a look at Table 3 to see one tutorials and overview papers, which follow other specific
example. Don’t forget to inform the p-value or specific rules.
parameters of the statistical tests cal- culated. Did you like it or not? is there something missing? did
you find typos or grammar mistakes? anyway, I am hun-
Running Time Experiments. This is the last and less im- gry for feedback and for new ideas to evolve this tuto-
portant experiment to be done in my opinion. With the tech- rial and help other machine learning researchers like you.
nology being evolved every time and with Moore’s law, Can you help me with this? send me a message at
what is considered slow today can be fast tomorrow. But anselmo.ferreira@gmail.com and let’s talk about it!
anyway, reviewers usually like this kind of experiment. All the figures used in this paper are from my pub-
So here, you calculate the running time of your approach lished works or works under review, so, I am not
and compare it against the baselines. You can use it for the citing them here. You can find my publications at
train- ing step, testing step, or both. Of course, it’s nice to http://www.ic.unicamp.br/ anselmoferreira. You can also
discuss why your approach is slower or faster, but I don’t find me on ResearchGate, GitHub, Google, Scopus and
believe that
Anselmo Ferreira | WritePaperDummies UNISI-DIISM | 13
Web of Science.

14 | UNISI- Anselmo Ferreira | WritePaperDummies

DIISM
ACKNOWLEDGEMENTS
Here you will thank everybody that helped in your work. This can be (i) some-
body that helped you in collecting datasets; (ii) somebody who sent you the source
code of their approach; (iii) somebody who helped you to make English proofread;
and more importantly (iv) the funding institutions that supported financially your re-
search. Don’t forget to inform the grant number of such support.

Bibliography
Different journals use different bibliography styles. Check
with your journal of interest about it.

Anselmo Ferreira | WritePaperDummies UNISI-DIISM | 9

View publication stats

16 | UNISI- Anselmo Ferreira | WritePaperDummies

DIISM