Interviews

"The deep learning applied in technology may determine a defendant's fate in a trial or a medical treatment"
 Gereziher Adhane

Photo: Gereziher Adhane

18/11/2021
Santiago Campillo
Gereziher Adhane, an Ethiopian doctoral student and researcher in the UOC

 

Adhane is a researcher at Ethiopia's Haramaya University who specializes in the analysis of artificial intelligence (AI) applied to deep learning and other technologies related to artificial neural networks. He is currently working on his doctoral thesis in the AI for Human Well-being research group (AIWELL, before SUNAI Lab), which is attached to the Faculty of Computer Science, Multimedia and Telecommunications and to the eHealth Center of the Universitat Oberta de Catalunya (UOC). He recently published, along with Mohammad Mahdi and David Masip, a research paper on the development of a technology that learns to identify mosquitos using a large volume of photographs taken and uploaded by volunteers to the Mosquito Alert platform.

What does your research activity consist of in the research group?

I am a PhD student in the research group under the supervision of David Masip. My research focuses on the application of deep learning models to computer vision and image processing for a variety of uses, including object detection and medical applications. My role is to create deep learning solutions and apply them to real-world problems.

You started your doctoral degree in 2019. What is the subject of your thesis?

I conduct research in the field of artificial intelligence. Although a lot has been published on the applications and uses of deep neural networks, when it is applied to critical areas such as medical apps, autonomous deriving and financial transactions, this technology presents numerous challenges. Some of the difficulties arise from not being able to estimate uncertainty and not being able to explain the decision-making process in deep learning. As a result, my research focuses on assessing uncertainty and decisions made based on these deep learning models.

You taught and carried out research at Haramaya University in Ethiopia. What was your line of research? Have you continued with it or changed its goals?

In Ethiopia I spent most of my time teaching courses related to artificial intelligence and carrying out research for a variety of applications. In addition to teaching and research I also worked as a supervisor for bachelor's degree students. Joining the research group laboratory did not substantially change the area of research I was involved in. I didn't find it hard to adapt to the working environment or the group I am now with.

What brought you to the UOC? What does the University offer to researchers in your specialist field?

When I decided to do a PhD programme I was looking for a suitable position in Europe. Luckily, I found the UOC's call for doctoral students and I decided to apply. I took a look at the University's website to find out more about its research areas and I found the research group. Once I was accepted, the UOC gave me a grant and additional resources for my research, which have helped my career.

You have published a scientific article on the usefulness of deep neural networks to identify mosquitos based on images with the Mosquito Alert platform. How did you become involved in this project?

Mosquito Alert is a platform that was launched in 2014 to monitor and control these disease carriers. The data gathering process is based on images of mosquitos and their breeding grounds uploaded by volunteers and subsequently inspected, validated and classified by a team of entomologists. This platform brings together members of the public, entomologists, public health authorities and mosquito control services to help reduce cases of diseases transmitted by mosquitos in Spain. At the UOC we took on the task of automatically annotating the images uploaded by the public to this platform.

Was it a complicated process? What was its most challenging aspect?

The main problem we found in the experiment was that some of the photos were not very clear. Many of them had background noise or mutilated parts of mosquitos. These photos are crucial to identify the species of mosquitos. Because the various species of mosquito are very similar in terms of physical appearance, poor-quality photographs can make it difficult to distinguish between closely related species. Thanks to advances made in machine learning, especially in deep neural networks, we were able to propose a deep convolutional neural network that could capture hidden correlations between images with similar morphological characteristics.

Can you elaborate on neural networks and how they can help society?

Neural networks, also known as artificial neural networks (ANNs), are a subset of machine learning that provides the base for deep learning techniques. Their name and form are based on the human brain and they replicate how biological neurons communicate with each other.

An ANN is made up of layers of nodes, which include an input layer, one or more hidden layers and an output layer. When a neural network receives input data, information patterns are learned within the network, activating the hidden layer units, which finally reach the output units to perform the classification work. 

When this calculation reaches a specific threshold, the unit is triggered, activating the units to which it is connected. For a neural network to learn, it has to receive feedback to reduce the gap between actual and predicted values. This procedure is repeated until we are satisfied with the performance of the neural network.

Deep neural networks are capable of performing a variety of crucial tasks, such as classification, prediction, grouping, medicine design, social medial filtration and natural language translation. We can also tackle more difficult scientific and engineering problems, such as advanced robotics, manufacturing and space navigation, for example.

It is a branch of science that seems to be highly specific. To what extent has the state of the art advanced in the use of neural networks?

ANNs are now being used for a wide range of real-word problems, normally in numerical paradigms for the approximation of universal functions due to their impressive qualities in terms of self-learning, fault tolerance, nonlinearity and progress in mapping from input to output.

Artificial neural networks can be used for everything from simple classification work to more complex optimization problems.

What are the main challenges facing machine learning, neural networks, algorithms and next-generation computing?

Deep neural networks, which belong to a broader family of machine learning technologies, can efficiently assess enormous volumes of data. In spite of this, there are still many obstacles hindering the implementation of deep learning models. Deep learning's lack of transparency, known as the "black box" problem, is a major concern in relation to the implementation of artificial intelligence (AI) models for commercial applications.

This is because we don't really know how AI reaches its conclusions. For example, the image recognition algorithms used in Google Photos, which classified black people as gorillas or the gender classification model that mainly classified black women as men and many other cases.

When deep learning is performing a trivial task, where an incorrect decision will have little to no effect, the black box problem is not serious. However, when the decision to be made concerns a defendant's fate in a court, a patient's medical treatment or pedestrian safety in self-driving vehicles, such mistakes may have catastrophic effects.

And what does this technology aspire to achieve?

Artificial intelligence helps us in our everyday lives. It aspires to provide machines that can interact with humans and perform activities like driving cars and working in factories or creating automated robots to work in high-risk environments such as nuclear power plants, mining and undersea gas exploration.

AI has been used in voice recognition and translation services, medical applications and weather forecasts and has a lot to offer in terms of improving the quality of our lives.

 

UOC R&I

The UOC's research and innovation (R&I) is helping overcome pressing challenges faced by global societies in the 21st century, by studying interactions between technology and human & social sciences with a specific focus on the network society, e-learning and e-health.

Over 500 researchers and 51 research groups work among the University's seven faculties and two research centres: the Internet Interdisciplinary Institute (IN3) and the eHealth Center (eHC).

The University also cultivates online learning innovations at its eLearn Center (eLC), as well as UOC community entrepreneurship and knowledge transfer via the Hubbik platform.

The United Nations' 2030 Agenda for Sustainable Development and open knowledge serve as strategic pillars for the UOC's teaching, research and innovation. More information: research.uoc.edu #UOC25years

Related links