An MDE Method for Improving Deep Learning Dataset Requirements
Engineering using Alloy and UML
Beno
ˆ
ıt Ries
a
, Nicolas Guelfi
b
and Benjamin Jahi
´
c
c
University of Luxembourg, Esch-sur-Alzette, Luxembourg
Keywords:
Model-driven Engineering, Software Engineering, Requirements Engineering, EMF, Sirius, Alloy.
Abstract:
Since the emergence of deep learning (DL) a decade ago, only few software engineering development methods
have been defined for systems based on this machine learning approach. Moreover, rare are the DL approaches
addressing specifically requirements engineering. In this paper, we define a model-driven engineering (MDE)
method based on traditional requirements engineering to improve datasets requirements engineering. Our
MDE method is composed of a process supported by tools to aid customers and analysts in eliciting, specifying
and validating dataset structural requirements for DL-based systems. Our model driven engineering approach
uses the UML semi-formal modeling language for the analysis of datasets structural requirements, and the
Alloy formal language for the requirements model execution based on our informal translational semantics.
The model executions results are then presented to the customer for improving the dataset validation activity.
Our approach aims at validating DL-based dataset structural requirements by modeling and instantiating their
datatypes. We illustrate our approach with a case study on the requirements engineering of the structure of a
dataset for classification of five-segments digits images.
1 INTRODUCTION
Deep Learning (DL) has emerged in the last decade
from artificial intelligence, dating from the Dart-
mouth conference in 1956, combined with the recent
emergence of Graphical Processing Units (GPUs).
These GPUs, providing a boost in computational
power, initiated the popularity of artificial intelligence
in everyday life applications, such as vocal personal
assistants, entertainment.
Deep learning techniques require large datasets ei-
ther for their training phase, in the case of supervised
learning neural networks, or for their learning phase
in the case of unsupervised learning networks. It is
common for real applications to have datasets as large
as tens of thousands of data. In one of our previ-
ous study (Jahic et al., 2019; Jahic et al., 2020), by
analysing the datasets and learning outcomes of the
training of neural networks, we discovered that many
issues were related to the poor specification of the
datasets’ structure.
Datasets are critical input artefacts necessary to
a
https://orcid.org/0000-0002-8680-2797
b
https://orcid.org/0000-0003-0785-3148
c
https://orcid.org/0000-0002-1120-1196
construct DL-based systems. As such they should be
precisely specified. Let’s take the following exam-
ple to illustrate our claim: a DL-based system aiming
to classify hand-written digits will have difficulty to
classify sevens without a bar if the training/learning
dataset comprises solely sevens with bars. With our
method proposed in this paper, we address this type of
issues by exploring the dataset requirements, before
the actual learning phase starts. Errors in datasets are
then detected earlier in the software engineering de-
velopment lifecycle, reducing consequently the cost
of dealing with these errors.
Software engineering appeared in 1968 (Naur and
Randell, 1969) as a solution to the so-called “soft-
ware crisis”. In the last 50 years, software applica-
tions have benefitted from research and development
effort in the area of software engineering. Unfortu-
nately DL-based systems are rarely defined system-
atically following a software engineering methodol-
ogy. DL approaches could take advantage of tradi-
tional software engineering methodologies. Typical
DL dataset requirements engineering is limited to an
activity whose output is informal and not based on
standard modeling languages neither grounded on for-
mal semantics. Datasets are also used in the context
of software test engineering. In this context, the se-
Ries, B., Guelfi, N. and Jahi
´
c, B.
An MDE Method for Improving Deep Learning Dataset Requirements Engineering using Alloy and UML.
DOI: 10.5220/0010216600410052
In Proceedings of the 9th International Conference on Model-Driven Engineering and Software Development (MODELSWARD 2021), pages 41-52
ISBN: 978-989-758-487-9
Copyright
c
2021 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
41
lection of relevant test datasets using model-driven
software engineering has shown to be an interesting
approach (Ries, 2009). It is only recently that inter-
est is growing in the research community to cross-
fertilize the aforementioned two areas (Hill et al.,
2016; Khomh et al., 2018; Vogelsang and Borg, 2019;
Burgue
˜
no et al., 2019).
In this paper, we provide a model-driven software
engineering (MDSE) method using semi-formal and
formal modeling. The dataset requirements engineer-
ing team is composed of the analysts, formal experts
and the customer. On the one hand, semi-formal mod-
eling allows one to describe concepts of the dataset
under development and to communicate them among
the heterogeneous stakeholders at the desired level
of abstraction. We use semi-formal modeling to de-
scribe the concepts required for the structure of the
datasets, i.e., the datatypes of interest.On the other
hand, formal modeling allows us to construct a pre-
cise description of the requirements on the structure
of the datasets. Formal interpretation tools like the
Alloy Analyzer and Kodkod provide instances of data
specification that satisfy the modeled requirements.
In this paper, we contribute to the introduction of
software engineering rigor in the requirements speci-
fication of datasets for DL-based systems. Section 2
presents our iterative process for dataset requirements
engineering with formal model execution. Section 3
illustrates our approach with a case study on the en-
gineering of the requirements for a five-segments dig-
its dataset with UML-compliant modeling and Alloy
models execution. Section 4 discusses some impor-
tant aspects. Section 5 positions our paper w.r.t. some
current related works. Finally, we conclude in Sec-
tion 6 and present some future works in Section 7.
2 AN ITERATIVE EXECUTABLE
DATASET REQUIREMENTS
MDE METHOD
2.1 Using Executable Models to
Improve the Engineering of
Datasets Requirements
In this paper, our iterative method focuses on the en-
gineering of datasets structural requirements. Con-
cretely, our focus is on the elicitation and modeling
of data types for the dataset under development. With
this method the DL scientist can follow a clear pro-
cess for dataset requirements engineering, supported
by tools, while benefitting from advances in estab-
lished software engineering methods.
Following traditional software engineering, our
method takes place in a typical requirements engi-
neering phase. This phase is usually (Sommerville,
2016) composed of the following activities: elici-
tation, specification, validation and evolution. Re-
quirements engineering is performed by an analyst
together with the customer of the system to be pro-
duced. Overall, requirements engineering is a crucial
phase. It is well-known that the workload put in pro-
ducing quality requirements reduces the amount of
workload put in the succeeding phases, i.e., design,
production, testing, deployment. Moreover, improv-
ing the requirements engineering phase eases the ear-
lier discovery of errors in the dataset under develop-
ment. Consequently, we make the hypothesis that im-
proving the engineering of the dataset requirements
phase will improve the dataset design and production
phases.
Our approach is based on executing dataset re-
quirements to improve their specifications. The so-
lution that we explore is to specify the datasets re-
quirements with an executable model, that we name
the dataset requirements concept model (DRCM).
DRCM is given an operational semantics as a trans-
lation to a formal language. Thanks to MDE tech-
niques, we generate a skeleton specification of the
formal DRCM (FDRCM). The formal analysts then
finalise the FDRCM. With this formal semantics, we
are able to interpret the dataset requirements concept
model. We name these formal interpretations: model
executions. A model execution is a set of datatype
specification instances. The resulting model execu-
tions are formatted in a way that should suit the cus-
tomer and the analyst such that they are able to decide
on the validation of the DRCM.
The objective of our method is to provide help
to the requirements analysts for the validation of the
datatypes that will structure the dataset. Typically, in
our context, these analysts are data scientists who are
responsible for the dataset engineering.
2.2 Iterative Dataset Requirements
Elicitation Process
2.2.1 Process Overview
Figure 1 shows the main elements of the business pro-
cess we propose in this paper using the BPMN (Ob-
ject Management Group, 2011) notation. Our process
is composed of: four activities denoted by dark-blue
horizontal rectangles; five artefacts denoted by light-
blue vertical rectangles with a folded top-right cor-
ner; one exclusive gateway denoted by a dark-blue di-
amond with a white X, this gateway allows numerous
MODELSWARD 2021 - 9th International Conference on Model-Driven Engineering and Software Development
42
iterations in our process until reaching the validation
of the dataset requirements concept model; one start
event denoted by a white circle; one end event denoted
by a dark blue-filled circle.
2.2.2 Model Dataset Structural Requirements
The first activity in our process is the modeling
of dataset structural requirements. Its objective is
to produce a Dataset Requirements Concept Model
(DRCM). As in a traditional requirements engineer-
ing phase, the output artifact of the requirements mod-
eling phase describes what are the constraints that the
system should satisfy. In our method, systems un-
der study are datasets in DL approaches. This model
should be compliant with the Metamodel given as in-
put artefact, see further Section 2.3.
In our approach, we concentrate on the structural
requirements of datasets for DL-based systems, i.e.,
on what should be the datatypes for the dataset un-
der development. The DRCM model describes the
requirements of the structure to be satisfied by the
dataset of the DL-based system under development.
The structural requirements may be described using
various modeling constructs, described in the further
Section 2.3.2.
During this activity, the main stakeholders are: on
the one hand the customer, and on the other hand an
analyst from the provider’s team. This activity may be
realised in a number of ways. Typically requirements
elicitation is performed through discussions among
the stakeholders and results in the elicitation of the
requirements model.
The content of the model under creation should
describe the concepts related to the required structure
of the dataset under development. More concretely
the data, their type and attributes as well as proper-
ties that they should satisfy, need to be elicited. This
activity is further described in Section 2.3.
2.2.3 Define a Formal Semantics
The second activity of our process is performed by a
formal language expert. It aims at producing an exe-
cutable formal requirements specification of the struc-
tural properties of the dataset under study. This activ-
ity takes as input the DRCM produced in the previous
activity. The format of the FDRCM is tool-dependent,
Section 2.3, below, gives insight on how we use the
Alloy language (Jackson, 2012) for this purpose.
Using formal techniques brings precision on the
definition of data structure and rigorous validation us-
ing mathematically defined semantics supported by
automated tools. This results in a higher quality of
requirements engineering.
2.2.4 Execute the Formal Specification
This activity consists in taking as input the FDRCM
specified earlier and to execute it. By executing, we
mean that the requirements should be fed to a formal
engine allowing one to query the formal model. The
aim of the queries is to provide enough data speci-
fication instances for the further validation activity.
It may then either comfort the analyst in the current
specification, or encourage the analyst to introduce
changes in the model. There are two kinds of queries
that we suggest to perform:
Exploration Queries. Such queries result in a set
of data specification instances. For instance, if
SevenEC is the name of the equivalence class that
should contain all images representing the digit
seven, then an example of exploration query is
to “find possible data specification instances of a
given DRCM satisfying the SevenEC equivalence
class”.
Verification Queries. Such queries check that the
dataset/data structure is coherent with some given
property. For instance, here is a sample veri-
fication query: “No data specification instance
should satisfy properties of more than one equiv-
alence class”.
This activity should be performed by formal ex-
perts together with analysts. On the one hand, the role
of the analyst is to identify the queries of best interest
to improve the dataset requirements, and on the other
hand the role of the formal expert is to make sure that
the queries are well-formalised.
2.2.5 Validate the Dataset Requirements
Concept Model
Its goal is the interpretation of the FDRCM execu-
tions, performed in the former activity, in order to
conclude whether these executions are satisfactory,
or not. If not satisfactory, then the requirements are
not validated, and the process proceeds to redoing an-
other iteration starting with the first activity Modeling
Dataset Structural Requirements. If the analysis of
the model execution results is satisfactory, then the
DRCM is validated and our process ends here. The
stakeholders may then continue with the succeeding
dataset engineering activities.
2.3 Syntax and Semantics of the DRCM
In this subsection, we give some details on the lan-
guage to be used for the modeling of the dataset struc-
tural requirements during the activities of our pro-
cess. We follow a typical modeling approach in three-
An MDE Method for Improving Deep Learning Dataset Requirements Engineering using Alloy and UML
43
Typical Structural
Dataset
Requirements
Engineering
Label
Names
Start
Event
(a)
not validated
Start
Model Dataset
Structural
Requirements
Define a
Formal
Semantics
Execute the
Formal
Model
validated
Metamodel Dataset
Requirements
Concept
Model
(DRCM)
Formal
Model
(FDRCM)
FDRCM
Executions
X
DRCM
Validate
DRCM
Event
(b)
Figure 1: A Typical Dataset Structural Requirements Engineering (a) in contrast with our Method (b).
levels, namely: meta-modeling, modeling and model
execution. Firstly, we define a metamodel that de-
scribes the types of modeling elements usable for the
modeling. Secondly, we present the scope of mod-
eling in our approach, i.e., dataset structural require-
ments. And lastly we discuss the semantics of DRCM
models allowing their consequent executions.
In our approach, we use the traditional standard
semi-formal modeling language UML (Object Man-
agement Group, 2017) to support the modeling ac-
tivity. It benefits from decades of industrial practice.
In our context of improving the DL requirements en-
gineering practices, we have chosen a small and re-
stricted set of UML concepts to ease its transfer to
non-Software Engineering practionners.
2.3.1 The DRCM Metamodel
A metamodel is a model defining the base elements
of a modeling language (Object Management Group,
2017). We defined the metamodel using the Eclipse
Modeling Framework, EMF (Steinberg et al., ), as an
ecore metamodel. Figure 2 is a diagram represent-
ing graphically our DRCM metamodel with the EMF
framework. In the following, we present shortly the
main concepts of a DRCM model through its meta-
model:
property is a boolean characteristic of the dataset,
data, or equivalence classes. It may be described
by its name, signature (ie name and variable
parameters), informal textual description, and/or
formal expression in Alloy.
Invariant Property: a property that is always true,
it is a constraint on a DRCM model that must be
satisfied.
Dataset represents the set of data under study. It
contains a set of properties and a set of invariants.
A given DRCM focuses on the description of a
single dataset.
Data: is a structured element. It contains fields
describing its structure. Fields are described as
typed variables, with a name and a type. The type
may either be a primitive type (Boolean, String,
Int) or another Data defined in the same DRCM
model. Data also contains a set of properties and
a set of invariants.
Data Equivalence Class: is a concept that repre-
sents a subset of the dataset’s data and is charac-
terised by a set of properties and invariant proper-
ties.
2.3.2 DRCM Modeling
In our approach, we use a subset of the base ele-
ments defined in the UML metamodel for class di-
agrams and follow the concrete graphical syntax of
MODELSWARD 2021 - 9th International Conference on Model-Driven Engineering and Software Development
44
Figure 2: The DRCM Ecore Metamodel.
UML class diagrams. Based on the metamodel pre-
sented above, we describe in the following the dataset
structural requirements concepts that may be modeled
in a DRCM, their graphical syntax and map them to
the corresponding UML elements. Each DRCM de-
scribes the requirements for a single dataset, a set of
data and a set of equivalence classes.
property is modeled as a UML Operation return-
ing a boolean value. A property predicate is de-
scribed as plain text between curly brackets {} af-
ter the operation signature. Note that the proper-
ties in our case study, as for instance, isInOneEC
of data equivalence class OneEC in Figure 4, are
described with Alloy language as a textual expres-
sion.
invariant property is modeled as a UML Opera-
tion returning a boolean value and prefixed by a
stereotype <<inv>>. As for properties, its expres-
sion may be described in curly brackets following
the operation’s signature.
Dataset: is modelled as a UML Class with
green header background. Its name is suf-
fixed with Dataset and the class has a stereo-
type <<dataset>>. Properties and invariants
may be described as UML Operations. See
FiveSegmentDigitsDataset class in Figure 4
for an illustration.
Data is modeled as a UML Class with blue header
background. Its name is suffixed with Data and
the class has a stereotype <<data>>. It possibly
contains UML Attributes to describe its fields and
boolean operations to describe its properties and
invariants. (e.g. FiveSegmentDigitData and
Segment classes in Figure 4).
Data equivalence classe is modeled as a UML
Class with red header background. Its name is
suffixed with EC and the class has a stereotype
<<equivalence-class>>. It contains boolean
operations to describe its properties and invari-
ants. (e.g. ZeroEC and OneEC classes in Figure 4).
2.3.3 Alloy Language-based Formal
Specification Syntax and Semantics
In this subsection, we define the following operational
semantics given as an informal translation from the
DRCM concepts to Alloy language (Jackson, 2012)
constructs.
DRCM classes are specified as Alloy signatures.
An MDE Method for Improving Deep Learning Dataset Requirements Engineering using Alloy and UML
45
DRCM metamodel classes are defined as abstract
signatures.
DRCM class attributes are defined as fields in the
related signature.
The semantics of operations of classes depend on
whether it has an <<inv>> stereotype or not.
DRCM class operations with an <<inv>>
stereotype are considered to be invariant prop-
erties. We define the semantics of such opera-
tions as Alloy signature facts.
DRCM class operations without stereotype are
considered to be properties to be checked in dif-
ferent contexts. As such, they are given the se-
mantics of Alloy predicates.
2.4 MDE Toolset for Our Approach
The toolset supporting our approach is composed of
four tools: a modeling language editor; a formal lan-
guage editor; a formal execution engine; and an ad-
hoc software application for the creation of synthetic
data based on their specification instantiations. The
source code of the toolset is available publicly (Ries,
2020).
2.4.1 DRCM Modeling Editor
To support the modeling of the DRCM requirements
model, we implemented a graphical modeling editor
based on the Sirius (Viyovi
´
c et al., 2014) framework.
The metamodeling concepts are defined in a .ecore
file based on Section 2.3.1. The concrete modeling
syntax is defined compliant with UML class diagram
syntax with the help of the Sirius framework.
When a DRCM is modeled, our toolset allows to
generate a FDRCM based on the information mod-
eled in the DRCM. This generation is implemented
with the XTend (Bettini, 2013) language well-suited
for such Model2Text model transformations.
2.4.2 FDRCM Textual Editor
Alloy Analyzer (Jackson, 2012) allows one to write
formal specifications in the Alloy language and of-
fers typical textual language editor features, such as
syntax highligthing and static analysis. We recom-
mend using the Alloy Analyzer to complete the for-
mal specification based on the FDRCM generated by
our Model2Text model transformation.
2.4.3 FDRCM Execution
Kodkod is a SAT-based constraint solver and is used
as a model finder (Torlak and Jackson, 2007) for spec-
ifications written in the Alloy language. It allows exe-
cuting Alloy models and provide instances satisfying
the given specification in the context of a given exe-
cution command. The Kodkod tool, now part of the
Alloy Analyzer tool, allows one to parse Alloy spec-
ification and provide possible executions, i.e., model
instances. This feature is particularly interesting in
our context as it allows us to perform queries on our
FDRCM, see Section 2.2.4.
2.4.4 Data Specification Requirements
Visualiser
The objective of this tool is to show a representation
of the data based on the specification instance gener-
ated by the Alloy formal engine. For instance, if the
specification describes a dataset of images, this tool
would create synthetic images based on the formal
data specifications instances created by Kodkod. In-
deed, and unfortunately, no generic tool would exist
that generates the representation based on the speci-
fication, as this is tightly coupled with the expected
concrete format of the data under specification. It is
thus necessary to develop an ad-hoc tool for this need.
In our toolset, we have implemented a Java program
that performs the FDRCM executions by calling the
Kodkod SAT-based constraint solver and interpreting
the model execution to create possible visual repre-
sentation of the data requirements specification in-
stances.
3 CASE STUDY: STRUCTURAL
REQUIREMENTS
ENGINEERING OF A
FIVE-SEGMENTS DIGITS
DATASET
The aim of this section is to provide a proof-of-
concept of our approach by conducting a step-by-step
instantiation of our process. This instantiation is per-
formed on the five-segments digits (FSD) case study
that we define in this paper. The DL-based system, for
which this dataset is defined, aims at recognising im-
ages of digits from 0 to 9. The case study is available
publicly (Ries, 2020).
3.1 Modeling the Five-segments Digits
DRCM
In this case study, a five-segments digit is charac-
terised by at most five segments, more precisely three
horizontal segments and two vertical segments, as
MODELSWARD 2021 - 9th International Conference on Model-Driven Engineering and Software Development
46
shown with the sample zero and nine digits in Fig-
ure 3.
0
9
hSeg1
hSeg2
hSeg3
(none)
vSeg2
vSeg1
hSeg1
hSeg2
hSeg3
vSeg2
vSeg1
Figure 3: Samples of Five-Segments Digits (a zero and a
nine).
In the first activity of our process, we elicit and
model the initial structural requirements for the FSD-
dataset. An extract of the resulting model of this ac-
tivity is shown in Figure 4. This activity is composed
of two tasks. The first task of this activity is to elicit
the main elements structuring the data in this dataset.
Here are some of the structural requirements that we
define:
Each data in the FSD-dataset shall be representing
one digit, from 0 to 9, using at most five segments.
The segments shall be line segments, either verti-
cal or horizontal.
There shall be at most three horizontal segments,
named hSeg1, hSeg2 and hSeg3 and at most two
vertical segments, named vSeg1, vSeg2.
The segments shall be characterised by their size,
and x-y position in an orthonormal 2D-space.
All data should be distinct, i.e., no two data having
the same segments
All horizontal segments should be distinct, i.e., no
two horizontal segments with the same x-y posi-
tion and the same size.
All vertical segments should also be distinct.
There is no empty data, i.e., no data without seg-
ment.
The second task of this activity is the elicitation
of the equivalence classes and their properties. In or-
der to ease the specification of equivalence classes’
properties, we defined a set of geometric property op-
erations: intersect, intersectT, isCornerTR, etc.
Let’s describe one of them: isCornerTR.
Property isCornerTR[seg1,seg2:Segment]
1
is
true when the segment seg1 ends
2
where the seg-
ment seg2 starts.
1
seg1 is assumed to be an horizontal segment and seg2
a vertical segment.
2
As a convention, the start, resp. the end, of an hori-
zontal segment is the point with the lowest x-value, resp.
the highest x-value. Similarly, the start, resp. the end, of a
vertical segment is the point with the highest y-value, resp.
lowest y-value.
Let’s describe informally the ZeroEC equivalence
class, modelled in Figure 4:
ZeroEC: a digit is a zero-digit if it has two hor-
izontal segments (hSeg1, hSeg2) and two ver-
tical segments (vSeg1, vSeg2) such that the
start of vSeg1 coincides with the start of hSeg1
(isCornerTL[hSeg1,vSeg1] is true), the end
of hSeg1 coincides with the start of vSeg2
(isCornerTR[hSeg1,vSeg2] is true), the end
of vSeg2 coincides with the end of hSeg2
(isCornerBR[hSeg2,vSeg2] is true) and the
start of hSeg2 coincides with the end of vSeg1
(isCornerBL [hSeg2,vSeg1] is true).
3.2 Specifying the Formal
Requirements for the Five-segments
Digits’ DRCM
We, taking the role of a formal expert, specify in Al-
loy each requirement elicited formerly in the DRCM
based on the FDRCM generated by our toolset. After
completion of the specification, the Alloy FDRCM is
as follows:
Firstly, the metamodel concepts:
3 abstract signatures for each of the 3 meta-
model concepts: Data, EquivalenceClass
and Dataset.
3 relations between: Data and
EquivalenceClass (named ec).
each metamodel properties, invariant or not, are
defined at the level of the model.
Then, the Alloy constructs for the specification of
the model concepts specific to the FSD case study:
1 signature FiveSegmentDigitDataset ex-
tending the Data signature.
10 signatures extending the
EquivalenceClass signature: ZeroEc,
OneEC, TwoEC, ThreeEC, FourEC, FiveEC,
SixEC, SevenEC, EightEC, NineEC.
each of the 10 signatures specify one predicate
characterising the equivalence class of a digit.
9 predicates for expressing the geometrical
boolean operations.
Finally, the Alloy invariants as 4 facts:
Specifying the metamodel’s
eachDataIsUnique invariant implied
writing two invariants at the model-level:
eachFSDDIsUnique and eachSegIsUnique.
Similarly for the metamodel’s invariant
noEmptyData, which resulted in the specifica-
tion of noEmptyFSDD and noEmptySegment.
An MDE Method for Improving Deep Learning Dataset Requirements Engineering using Alloy and UML
47
Figure 4: Case Study: Extract of the Dataset Requirements Concept Model Resulting from the First Iteration of Our Process.
3.3 Executing the Formal Specifications
In this activity, we define two Alloy commands to run
two different executions of the DRCM model.
The objective of the first execution is to explore
the possible data specification instances. We spec-
ify an Alloy command to instantiate all digits of size
5x5. Figure 5a shows one of the instance generated
from the execution of this command. In this figure,
the model execution corresponds to a specification of
a digit with 3 segments: one horizontal segment start-
ing in position (2,2) of size 1, one vertical segment
starting in position (2,1) of size 1 and a second verti-
cal segment starting in position (3,2) of size 2. This
instance satisfies the property of the FourEC equiva-
lence class.
Thanks to the Alloy Analyzer tool API and the
Java program that we develop to support this ap-
proach, the generated Alloy specification is inter-
preted and a file is created to represent the five-
segment digit specification instance as an image, see
Figure 5b.
In this execution, we generate, more than 400 dif-
ferent digit specifications. Figure 6 shows a visual
interpretation of an extract of these generated digits.
We also perform a second model execution with
the objective to check that there is no data satisfy-
ing the properties of more than one equivalence class.
It returns a non-empty set of digits that satisfy two
equivalence classes: OneEC and SevenEC. Figure 7
shows the interpreted results of this second Alloy
model execution.
MODELSWARD 2021 - 9th International Conference on Model-Driven Engineering and Software Development
48
(a)
(b)
Figure 5: An Alloy model execution (a) and its visual rep-
resentation (b).
Figure 6: Extract of the interpreted digits instantiated by the
exploration query.
3.4 Validating the DRCM
In this last step of our process, we analyse the
FDRCM execution in order to evaluate if the DRCM
is validated, or not. Here are the three points conclud-
ing the analysis:
1. We are satisfied by the model executions con-
cerning the characterisation of the zero-digit data
structure, thus we validate the DRCM related to
the zero-digit equivalence class.
2. The verification execution is not satisfactory. In-
deed, in this case study, we expect that the relation
between digit instances and equivalence class is
unambiguous by relating a digit data with exactly
one equivalence class. Thus this execution ex-
Figure 7: Digits corresponding to specifications in more
than one equivalence class.
hibits an issue in our DRCM, namely the fact that
several specification instances satisfy the proper-
ties of both one-digit’s equivalence class (OneEC)
and seven-digit’s equivalence class (SevenEC).
3. We wish to improve the DRCM such that nine-
digits without the bottom horizontal segment
should also be part of the dataset requirements.
In the end, we conclude that the DRCM of this
first iteration is not validated due to the two identified
issues in points 2. and 3. above. Thus there is a need
to perform a second iteration of the process.
3.5 Second Iteration of the Process
3.5.1 Updating the DRCM and the FDRCM
In the second iteration, we update the DRCM and the
FDRCM. We introduce modifications on the property
descriptions of:
the NineEC in which we state that the third hori-
zontal segment is optional with a logical implies
statement.
the OneEC. There should be now either zero hori-
zontal segments or two horizontal segments, but
never a single horizontal segment such that the
ones are not mistaken for sevens.
3.5.2 Executing the Updated FDRCM and
Validation of the Updated DRCM
We define a first Alloy command and run it to explore
the new nine-digits specification. Figure 9 shows
an extract of some of the new nine digits result-
ing from the change in the DRCM’s model property
isInNineEC of the NineEC class. A second command
is defined and we execute it to confirm that there is
no digit satisfying the properties of two equivalence
classes at the same time. In particular, the one-digits
and the seven-digits are now partitioned. Thus in the
end of this second iteration, we validate the DRCM
and end the process.
Consequently, we may then go on with the further
activity in the dataset development process, namely
dataset construction, which is out of the scope in this
paper. Among possible activities, this could typically
involve: data labelling, data preparation, data synthe-
sis, etc.
As a conclusion, we improved the quality of the
dataset by using our method and tool. Firstly by pro-
viding a model of the dataset structural requirements.
Secondly our method allowed us to detect the issue
with ambiguous ones and sevens, as well as allowing
us to explore a different possible kind of nines. Our
An MDE Method for Improving Deep Learning Dataset Requirements Engineering using Alloy and UML
49
Figure 8: DRCM Changes in Classes OneEC and NineEC.
Figure 9: Explored Model Executions of Nines Without 3rd
Horizontal Segment.
method helped detect these issues earlier in the DL-
based system development process.
4 DISCUSSION
In this section, we discuss shortly about some impor-
tant aspects of this paper. The first aspect is the scal-
ability of our method. In this paper, we present our
model-driven engineering method for requirements of
datasets and we validate it experimentally by apply-
ing it to the five-segments digits case study. Our
case study is of rather low combinatory complexity.
Due to this low complexity, we could specify its re-
quirements models: DRCM and FDRCM at a low-
level of abstraction, with precise details on the data
structure, i.e., horizontal/vertical segments and their
size/position. We could also execute all model exe-
cutions due to the low number of combinations, i.e.,
around 400 possible data instances. Let’s note that
it is rarely possible to be able to perform all possi-
ble model executions. So first of all, it is most im-
portant that there should be enough model executions
to either comfort the analyst and the customers that
their requirements model is valid; or to exhibit some
requirements to be improved thanks to our so-called
exploration or validation queries. Secondly, require-
ments models allow to describe conceptual domains
at a chosen level of abstraction. Thus, as for any mod-
eling methods, it must be decided what is the right
level of abstraction adapted to the project under devel-
opment. Indeed, when using our method for a larger-
scale case study, the level of abstraction should con-
sequently be higher in order to be tractable. This is
especially true for formal methods whose interest is
not in being applicable for a full real system but in
its contribution to increase the quality of the system
under development.
The second aspect that we would like to discuss
is the toolset limitation. The presented toolset is a
prototype implementation, intended to make a proof-
of-concept in terms of tool-support for our method.
Some parts of the toolset must indeed be improved be-
fore being transfered to an industrial context, or used
at a larger scale. For instance, some limitations in-
clude, the fact that the different textual editors are the
default ones provided by EMF, thus there is no syntax
checking, nor syntax highlighting on the textual ex-
pressions written in Alloy, nor rich text editing on the
properties and invariants descriptions.
5 RELATED WORK
Only few works are available on requirements engi-
neering for DL-based systems. The first one by Hill
et al. (Hill et al., 2016) presents a process comprising
nine stages for a machine learning workflow, but with-
out going into the details of the first stage, model re-
quirements. The second one by Amershi et al. (Amer-
shi et al., 2019) present a general SE process for ML-
based systems. Their process includes a requirements
engineering phase that considers the system in the
large without giving particular concerns about the re-
quirements of the dataset.
Industry also contributes to this area of ML-based
systems development. As for instance, Google de-
scribes a ML Workflow (Google, 2020), but the first
activity is Source and Prepare your data and the
requirements engineering is basically skipped. Mi-
crosoft defines the TDSP (TDSP, 2020; Mathew et al.,
2018) process, standing for Team Data Science Pro-
cesss. This process is composed of a lifecycle start-
ing with a Problem Definition phase, with an activity
Data characteristics questions. Unfortunately, these
two related works lack support for semi-formal and
formal modeling of dataset structural requirements.
In the MDE community, common interest in MDE
MODELSWARD 2021 - 9th International Conference on Model-Driven Engineering and Software Development
50
and AI is also quite recent, a first workshop on AI and
MDE was held in 2019 (Burgue
˜
no et al., 2019).
Regarding model execution, a large number of re-
lated works are available, published in the last two
decades in the literature, around the formalisation
and execution of UML models (Object Management
Group, 2018; Shah et al., 2009; Mellor et al., 2002),
but none of these approaches tackle specifically the
modeling of datasets requirements for DL-based sys-
tems.
Lastly, a number of approaches, mainly in the
AI community, e.g. DeepXplore (Pei et al., 2017),
Scenic (Fremont et al., 2019), DeepTest (Tian et al.,
2018), CARLA (Dosovitskiy et al., 2017), have been
published on designing and generating datasets for
DL-based systems. Whereas in our paper, we provide
an iterative method involving the customer and SE/DS
analysts focused on the production of a requirements
specification model, i.e., our method is not about de-
signing, nor producing datasets.
6 CONCLUSION
In this paper, we contributed to the introduction
of software engineering rigor in the engineering of
datasets for DL-based systems. We presented our
model-driven engineering method composed of a pro-
cess defined with BPMN and supported by a toolset.
Our Dataset Requirements Concept Model (DRCM)
is modeled in compliance with UML syntax. We pre-
sented a DRCM graphical editor implemented EMF
and Sirius that allows creating diagrams to ease the
requirements elicitation discussions. The DRCM
model is executable with the Alloy Analyzer tool,
thanks to our informal definition of the DRCM opera-
tional semantics as a translation from UML to Alloy.
Our Xtend implementation allows us to generate a
FDRCM skeleton code in Alloy. Lastly, we presented
an instantiation of our process in a case study related
to the requirements engineering of a ve-segments
digits dataset and discuss some current issues.
7 FUTURE WORK
In this paper, we present a method for improving DL
dataset requirements engineering, which has raised a
number of interesting future works.
In order to increase the MDE dimension of our
approach, in our method, the definition of the transla-
tional semantics of DRCM models will firstly be im-
proved using an Alloy metamodel. Together with the
definition of Model-2-Model transformations from
DRCM concepts to Alloy concepts. Secondly, we
plan to define a textual domain-specific language
(DSL) to allow specifying the DRCM textually. The
textual and graphical DRCM will be synchronised
such that changes in diagrams update the textual
counterpart and vice-versa.
Some other future works are related to the rigor-
ous validation of our method. For this, we plan to con-
duct additional experimentations, with well-known
examples like MNIST (Yann et al., 2018). These ex-
perimentations will serve for comparative evaluation
of our method against existing methods on common
case studies. Moreover, we will apply our method on
a larger-case study, in the area of earth’s ecosystems
health monitoring. This latter case study will investi-
gate scalability, practicality, and possibly adaptation,
of our method for potential transfer to larger exam-
ples.
Lastly, some final future works are related to the
software engineering aspects of our approach. We
will improve our approach in two ways. Firstly, we
will broaden the work on the dataset requirements
engineering phase by including other types of func-
tional and non-functional requirements. Moreover
this phase should be placed in the more general con-
text of DL-based software requirements. How do
the dataset requirements involve (or are involved in)
the DL-based software requirements ? Secondly, we
will concentrate on the definiton of the succeeding
dataset engineering lifecycle phases, e.g., dataset de-
sign, dataset production, dataset assessment, dataset
deployment.
REFERENCES
Amershi, S., Begel, A., Bird, C., DeLine, R., Gall, H., Ka-
mar, E., Nagappan, N., Nushi, B., and Zimmermann,
T. (2019). Software Engineering for Machine Learn-
ing: A Case Study. In 2019 IEEE/ACM 41st Inter-
national Conference on Software Engineering: Soft-
ware Engineering in Practice (ICSE-SEIP), Montreal,
Canada.
Bettini, L. (2013). Implementing Domain-Specific Lan-
guages with Xtext and Xtend: Learn How to Imple-
ment a DSL with Xtext and Xtend Using Easy-to-
Understand Examples and Best Practices. Commu-
nity Experience Distilled. Packt Publ, Birmingham.
Burgue
˜
no, L., Burdusel, A., G
´
erard, S., and Wimmer, M.
(2019). Preface to MDE Intelligence: 1st Workshop
on Artificial Intelligence and Model-Driven Engineer-
ing. In ACM/IEEE 22nd International Conference on
Model Driven Engineering Languages and Systems
Companion.
Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A., and
An MDE Method for Improving Deep Learning Dataset Requirements Engineering using Alloy and UML
51
Koltun, V. (2017). CARLA: An Open Urban Driving
Simulator. arXiv:1711.03938 [cs].
Fremont, D. J., Dreossi, T., Ghosh, S., Yue, X.,
Sangiovanni-Vincentelli, A. L., and Seshia, S. A.
(2019). Scenic: A Language for Scenario Specifica-
tion and Scene Generation. Proceedings of the 40th
ACM SIGPLAN Conference on Programming Lan-
guage Design and Implementation - PLDI 2019, pages
63–78.
Google (Last visited on Sept. 2020). Machine
learning workflow. https://cloud.google.com/ai-
platform/docs/ml-solutions-overview.
Hill, C., Bellamy, R., Erickson, T., and Burnett, M. (2016).
Trials and tribulations of developers of intelligent
systems: A field study. In 2016 IEEE Symposium
on Visual Languages and Human-Centric Computing
(VL/HCC), pages 162–170, Cambridge. IEEE.
Jackson, D. (2012). Software Abstractions: Logic, Lan-
guage, and Analysis. The MIT Press.
Jahic, B., Guelfi, N., and Ries, B. (2019). Software En-
gineering for Dataset Augmentation using Genera-
tive Adversarial Networks. In Proceedings of the
10th IEEE International Conference on Software En-
gineering and Service Science (ICSESS 2019), Bei-
jing, China. IEEE.
Jahic, B., Guelfi, N., and Ries, B. (2020). Specifying Key-
Properties to Improve the Recognition Skills of Neural
Networks. In Proc. of the 2020 European Symposium
on Software Engineering, Roma, Italy. ACM.
Khomh, F., Adams, B., Cheng, J., Fokaefs, M., and Anto-
niol, G. (2018). Software Engineering for Machine-
Learning Applications: The Road Ahead. IEEE Soft-
ware, 35(5):81–84.
Mathew, S., Danielle, D., and Tok, W. H. (2018). Deep
Learning with Azure - Building and Deploying Arti-
ficial Intelligence Solutions on the Microsoft AI Plat-
form. Apress.
Mellor, S. J., Balcer, M., and Jacoboson, I. (2002). Exe-
cutable UML: A Foundation for Model-Driven Archi-
tectures. Addison-Wesley Longman Publishing Co.,
Inc., USA.
Naur, P. and Randell, B. (1969). Software Engineering
Report of a conference sponsored by the NATO Sci-
ence Committee Garmisch Germany 7th-11th October
1968.
Object Management Group (2011). Business Process
Model and Notation (BPMN) v2.0. OMG Standard
Full Specification formal/2011-01-03.
Object Management Group (2017). Unified Modeling Lan-
guage: Superstructure (UML), v. 2.5.1. OMG Stan-
dard Full Specification formal/17-12-05.
Object Management Group (2018). Semantics of a Founda-
tional Subset for Executable UML Models (fUML), v.
1.4. OMG Standard Full Specification formal/18-12-
01.
Pei, K., Cao, Y., Yang, J., and Jana, S. (2017). DeepXplore:
Automated Whitebox Testing of Deep Learning Sys-
tems. Proceedings of the 26th Symposium on Operat-
ing Systems Principles.
Ries, B. (2009). SESAME - A Model-Driven Process for
the Test Selection of Small-Size Safety-Related Em-
bedded Software. PhD thesis, University of Luxem-
bourg, Luxembourg.
Ries, B. (2020). DRCM Editor and FSDD Case Study.
https://doi.org/10.5281/zenodo.4020938.
Shah, S. M. A., Anastasakis, K., and Bordbar, B. (2009).
From UML to Alloy and back again. In Proceedings
of the 6th International Workshop on Model-Driven
Engineering, Verification and Validation - MoDeVVa
’09, pages 1–10, Denver, Colorado. ACM Press.
Sommerville, I. (2016). Software Engineering. Pearson,
tenth edition edition.
Steinberg, D., Budinsky, F., Paternostro, M., and Merks,
E. Eclipse Modeling Framework Second Edition.
page 14.
TDSP (Last visited on Sept. 2020). The Team Data
Science Process. https://docs.microsoft.com/en-
us/azure/machine-learning/team-data-science-
process/.
Tian, Y., Pei, K., Jana, S., and Ray, B. (2018). DeepTest:
Automated Testing of Deep-Neural-Network-driven
Autonomous Cars.
Torlak, E. and Jackson, D. (2007). Kodkod: A Relational
Model Finder. In Grumberg, O. and Huth, M., editors,
Tools and Algorithms for the Construction and Analy-
sis of Systems, volume 4424, pages 632–647. Springer
Berlin Heidelberg, Berlin, Heidelberg.
Viyovi
´
c, V., Maksimovi
´
c, M., and Perisi
´
c, B. (2014). Sir-
ius: A rapid development of DSM graphical editor.
In IEEE 18th International Conference on Intelligent
Engineering Systems INES 2014.
Vogelsang, A. and Borg, M. (2019). Requirements En-
gineering for Machine Learning: Perspectives from
Data Scientists. arXiv:1908.04674 [cs].
Yann, L. C., Corinna, C., and Christopher, B. (2018). THE
MNIST DATABASE of handwritten digits.
MODELSWARD 2021 - 9th International Conference on Model-Driven Engineering and Software Development
52