Keynote Speakers

Friedhelm Schwenker. Deep learning architectures in Emotional Artificial Intelligence

Yukihiro Kamiya. A New Signal Processing Approach Suitable for IoT and Challenges for Its Commercialization

Vladimir Stanovov. Differential Evolution: trends, challenges, ways to visualize, understand and improve

Mikko Kolehmainen, Kristina Brester. The next step after AI – towards Digital Twins with examples from the energy sector

Deep learning architectures in Emotional Artificial Intelligence

Abstract. In this talk we introduce automatic affect recognition based on multimodal sensory data. In the first part we introduce learning paradigms, classifiers, and information fusion architectures for the challenging task of multimodal affect recognition, particularly the focus is on classifier fusion, semi-supervised learning, and the extraction of discriminative features from audio, video, and bio-signals. We introduce our ATLAS framework which is an efficient tool to analyze and annotate multimodal sensory data streams. The second part is evoted to the presentation of case studies on different affective data sets (University Ulm Multimodal Affective Corpus UUlmAC, SenseEmotion) that have been recently carried out in our lab.

A New Signal Processing Approach Suitable for IoT and Challenges for Its Commercialization

Abstract. It is said that the world economy is now driven by startups, not by legacy large-scale companies. ICT Technopolis Research Institute (ITRI) was established at Aichi Prefectural University in April 2021, aiming to support the creation of startup ecosystem in Aichi prefecture. In this talk, first, an overview of ITRI and its objective are introduced including entrepreneurship education programs for students and support programs for professors to commercialize their technologies. Second, we introduce a challenge for the commercialization of our technology based on a new signal processing approach termed as Accumulation for real-time serial-to-parallel converter (ARS) which is suitable for applications on the Internet of Things (IoT).

Differential Evolution: trends, challenges, ways to visualize, understand and improve

Abstract. The Differential Evolution (DE) is the prize-winning continuous optimization method, which is relatively simple to implement and apply, but the reasons for its exceptional performance are often unclear. This talk will be focused on revealing the inner mechanisms allowing DE to be efficient, and will use visualization with Expected Fitness Improvement for that purpose. The main problem of DE – the parameter tuning – will be also discussed, including most efficient adaptation techniques, their strengths and weaknesses, as well as some ideas for improvement. In particular, the greediness of parameter adaptation will be considered, and the biased adaptation to overcome it. Finally, the computational approach for automatic generation of parameter adaptation techniques with Genetic Programming will be presented.

The next step after AI – towards Digital Twins with examples from the energy sector

Abstract. As the wave of Artificial Intelligence (AI) rises and reaches the mainstream computing, it is time to look forward and think about the next wave after this development. Firstly, we will have a look at the main characteristics of the current AI wave and its benefits. The focus is then shifted to the weak points of it showing what we need to do better. Another hot topic rising is that of Digital Twins (DTs). What is their relation to AI? What are DTs actually? How could we combine these two technologies to yield better solutions?

Next, when AI and DTs have been discussed at the conceptional level, we will proceed to show examples with the energy sector, where the latest innovations already make the forth industrial revolution happen. Being implemented for individual houses, DTs constitute grounds for modeling the whole network, which opens new opportunities for dynamic load control, consumption optimization, network resilience improvement, etc.

On the way to building a network-level DT, questions inevitably occur, and some of them have been studied with a couple of example projects at the University of Eastern Finland during the last years. In particular, we investigated how to separate and predict heating loads at the single house level. Then, we compared physics-based and data-driven models in predicting solar photovoltaic output. Lastly, we learned how to forecast weather-related faults in the energy network. These examples will be introduced in the presentation, and finally, recommendations for future research directions will be given.