Skip to main content

NeCOL: An Innovative Methodology for Building Better Deep Learning Tools for Real Word Applications

Periodic Reporting for period 2 - NeCOL (NeCOL: An Innovative Methodology for Building Better Deep Learning Tools for Real Word Applications)

Reporting period: 2020-09-10 to 2021-09-09

Over the last years, the vast increase of digital data and the access to powerful computation resources have stimulated the fourth industrial revolution through the application of Artificial Intelligence (AI), specifically the application of deep neural networks (DNN), a.k.a. deep learning (DL). In spite of DL is being used to design breakthrough applications, the efficient design and training of DNNs is still an open problem and there is room for improvement. In this context, evolutionary computation (EC), which has been successfully used to solve hard-to-solve real-world optimization problems, emerges as an excellent tool to address DNN optimization. Here, we focus on recurrent neural networks (RNNs) and generative adversarial networks (GANs)

Thus, this project has the relevant scientific goal of the development of a cutting edge DL methodology based on deep evolutionary neural networks (DENN), i.e. the combination of EC, specifically co-evolutionary algorithms (CEAs) and evolutionary algorithms (EAs), and DNN. Furthermore, we apply DENN to address applications of importance in present societies, such as cybersecurity and Smart Cities (SC). Hence the name of the project: Neural CO-evolutionary Learning or NeCOL. NeCOL will comprise a framework based on CEAs to deal with DL
The work carried out during the first two years of NeCOL is grouped by work packages (WP).
WP1. RNNs optimization problem definition and WP2. Design and implementation of NeCOL by using CEAs
We defined a new methodology based on EC that allows the efficient selection of the DL training parameters while reducing the computational cost of this process, without needing to train the networks to evaluate them. In turn, we published several outstanding studies by applying parallel CEAs to train GANs. This piece of research is included over the name of Lipizzaner.
Besides the Lipizzaner framework, the main outcomes of these WPs at the moment of the writing report are nine scientific papers: three published and one under review in high impact journals, a book chapter, and four presented and published in high ranked conferences.
WP3. NeCOL use case I. Cybersecurity
The researcher reviewed the literature about applying CEAs to address cybersecurity problems and publish a review paper in the special issue of the 20th anniversary of the Genetic Programming and Evolvable Machine journal.
WP4. NeCOL use case II. Smart Cities
We addressed three different types of SC problems of high value and relevance in EU society: smart mobility, smart waste management, and smart energy. In turn, we include in this WP the studies of applying NeCOL to address one of the most challenging problems currently facing our society, COVID-19. The publications produced during this period related to this WP are 13: five conferences and eight journals; which cover a broad number of applications and methodologies.
a) Smart mobility: Regarding smart mobility, the researcher has opened three different branches of work related to NeCOL: the evaluation of mobility policies, the application of DL to improve data-driven pollution forecasting, and proposing a communication protocol to provide Intelligent Transportation Systems (ITS).
b) Smart waste management: We proposed the use of RNNs to predict the filling level of the waste containers. Besides, we addressed the efficient localization of the waste collection points (containers and bins).
c) Smart energy: The researcher applied CEAs to provide electricity consumption schedules taking into account user satisfaction and service cost.
d) COVID-19: We studied the use of the power of generative models trained by using NeCOL to synthesize X-Ray COVID-19 images to be used by computational intelligence methods.
WP6. Training and career development
This WP entails the activities aimed at advancing in the PCDP and at a correct development of NeCOL. These activities include attending and participating in seminars, meetings, workshops, etc.
WP7. Dissemination and public engagement
a) Research publications:
During this period of time, the outcomes of the research carried out under NeCOL project have been accepted and published as papers in: 11 journals, one book chapter, and nine international conferences. In addition, we have submitted three different works to be published in high impact journals and conferences, which nowadays are under review.
b) Complementary research outreach:
In addition to the publications, the researcher has participated in four different workshops and poster sessions and seven invited talks. In turn, he published about the actionon the Internet (project websites ( and and social media. It is noticeable his participation: first, in the MSCA Artificial Intelligence Cluster (December 2019) organized by the Research Executive Agency (REA); and second, in the Falling Walls Lab Marie Skłodowska-Curie Actions 2020 ( in which he was a finalist with a research idea entitled Breaking the Wall of Actionable Climate Intelligence. In turn, we remark his presentation, named Smart City Tools to Evaluate the Impact of Car Restrictions Policies in Urban Areas: Madrid Central Case, during the First ECUSA Boston Symposium, which was awarded the Best Oral Presentation award.
Moreover, the researcher worked for reaching the goal of acting as a Marie Curie Ambassador in order to reach a wider audience with the message about the high value of the MSCA fellowship career progression of the excellent scientists in Europe. Thus, he joined the MCAA North America Chapter. First as an associated, he was the fellow of the month. Later he decided to be part of the management board.
The research outcomes of this action can be mainly grouped in two different categories: algorithmic progress on DENN and application of NeCOL to address real-world high impact problems.

Thus, during this project, we have been able to define a new methodology based on EC that allows the efficient selection of the training parameters while reducing the computational cost of this process. The researcher has been able to contribute with several outstanding research studies by applying parallel CEAs to train GANs.

The main SC real-world problems proposed to be addressed was smart mobility, mart waste management, and smart energy. Among these research work, we include the application of efficient DL models to address prediction and modeling, the application of CEAs to optimize maximize quality of service and minimize resources, and the use of statistics and DL to evaluate policies and measures. In turn, we propose a data augmentation problem to address the lack of X-Ray COVID-19 images to train computational intelligence and learning methods to assist physicians in the process of diagnosing diseases.