The International Work-Conference on Artificial Neural Networks (IWANN) is a biennial meeting that, since the first edition in 1991, seeks to provide a discussion forum about systems inspired by nature, focusing, in particular, on trending topics such as Deep Learning and Explainability. Besides conventional talks and special sessions, IWANN 2023 has included complementary activities, such as a tutorial on Deep Learning Interpretability, and a round table to debate the timely topic of Generative Artificial Intelligence.

This 17th edition has been held in Azores, Portugal, enriched with the plenary talks by the outstanding researchers Amaury Lendasse (University of Houston, USA) and Alberto Bosio (École Centrale de Lyon, France). More than one hundred papers were been accepted, submitted by authors from over 40 different countries, comprising five continents. Also the strong participation of young researchers is particularly rewarding.

The warm atmosphere and relaxed mindset are distinguishing features of IWANN, which often encourage fruitful discussion and further collaboration.

Miguel Atencia & Gonzalo Joya. 

*****

Janine Strotherm PhD candidate at Bielefeld University, Germany. 

Can you tell us a bit about yourself and your research interests?

I am Janine Strotherm, a PhD student within the Machine Learning Group of the Technical Faculty at Bielefeld University. Originally, I studied Maths and Physics, but got interested into the applications of Mathematics within the world of Machine Learning (ML). Visiting some of the related lectures, I met my now-supervisor Prof. Dr. Barbara Hammer and joined the project “Water Futures” funded by the European Research Council (ERC) (https://waterfutures.eu/) after finishing my Masters. Within this project, we work on methods in which ML can support tasks within Water Distribution Networks, such as leakage detection. In my case, specifically, I focus on how to introduce the relevant topic of fairness into the domain of ML-assisted Water Distribution Network tasks.

What first interested you in this field of research?

Guaranteeing access to clean drinking water, Water Distribution Networks display critical infrastructure. At the same time, studies show that water scarcity will be a significant problem in the future. Therefore, if a ML-system makes decisions related to such infrastructure, fairness of such decisions is essential. However, while there exists a lot of research within the domain of water-related ML-systems already, to the best of our knowledge, none of them considered fair ML so far. Also in other human-affecting and ML-based systems, fairness is a highly relevant topic. As the usage of ML in our daily lives increases, this topic seemed quite appealing to me.

You won the Best Poster award at IWANN 2023, can you briefly explain the research you presented?

The paper which I presented at IWANN 2023 is called “Fairness-Enhancing Ensemble Classification in Water Distribution Networks” and was motivated by the question of how to introduce fairness into mechanisms we already use within the domain of Water Distribution Networks. More specifically, the idea is to make a ML-based leakage detection model, which we proved to not be fair, fair in the sense that the location where the leakage is present has no influence on the quality of the leakage detection. If that was be the case, instead, people living at places where the model detects leakages significantly worse compared to other regions would be discriminated by only the location they live at due to a higher risk of water scarcity, which is considered as unfair. The paper introduces fairness-enhancing algorithms that allow to train models that are fair in this sense.

What are you next steps?

My current work was a first step on the one hand to introduce the notion of fairness within the domain of Water Distribution Networks and on the other hand to prove that adaptations of well known fairness-enhancing algorithms work in such setting. However, the Water Distribution Network used is a first-step toy network and the leakage detection model is based on rather simple linear ensemble methods. Obvious next steps will be to test the solutions on more realistic networks using more powerful leakage detection models, such as Graph Neural Networks.
Moreover, as soon as suitable data will be available, the fairness question will be even more interesting in more relatable fields such as water quality control, network planning or pricing policies using ML.

El Mahdi Mercha PhD candidate at ENSIAS, Mohammed V University, Morocco. 

Can you tell us a bit about yourself and your research interests?

I am a fourth-year PhD candidate at ENSIAS. My journey began with a focus on applied mathematics. However, driven by the immense impact of machine learning and deep learning in various domains, I decided to delve into these advancements. Presently, my research revolves around advanced deep learning techniques and graph neural networks for multilingual sentiment analysis including dialects. Broadly, my research interests include information retrieval, computer vision and adversarial machine learning..

What first interested you in this field of research?

Despite extensive research in sentiment analysis, the predominant focus has been on analyzing sentiment in specific languages. However, given that web users tend to express themselves in their native languages, the content available for analysis is multilingual in nature. Consequently, the need to support sentiment analysis across languages has become increasingly crucial. In light of this, and considering the effectiveness of graph neural networks, I find it particularly interesting to explore the development of multilingual sentiment analysis approaches based on graph neural networks.

You won the Best Presentation award at IWANN 2023, can you briefly explain the research you presented?

The research work I presented at IWANN 2023 was entitled ”SlideGCN: slightly deep graph convolutional network for multilingual sentiment analysis”. In this work, we propose an approach for multilingual sentiment analysis using a graph convolutional network. We construct a single heterogeneous text graph to model the entire multilingual corpus. The words and documents serve as nodes in the constructed graph. The relationships between these nodes are built based on semantic, sequential, and statistical information. Then, we propose SlideGCN (SLIghtly DEep Graph Convolutional Network), to model the graph. The task of multilingual sentiment analysis is then considered as a node classification problem. The proposed approach can learn insightful representations for words and documents, and it achieves strong classification performance without using external sentiment information.

What are you next steps?

I intend to generalize our proposed approach to inductive settings. Also, I plan to work on the optimization of the memory consumption in large corpora. In addition, I will conduct more extensive experiments on other new language combinations with new datasets to investigate the robustness of the proposed approach against language variation. Furthermore, I plan to integrate our approach into real big data systems.