Keynote Talks for ACSS 2025

Dr. Piotr Porwik

Institute of Computer Science Faculty of Science and Technology University of Silesia, Poland



Noise-tolerant method for detecting feature drift in a data stream

Drift detection methods play a crucial role in identifying changes within data streams. Conventional drift detection techniques often operate assuming input data is free from noise. However, in practical applications, data streams are frequently affected by various types of noise, including class or attribute noise. These phenomena occur, for example, in medical sensors monitoring the patient's vital signs in ICUs. In the speech, we show in detail a novel feature-based drift detection method, which employs a two-stage process involving feature selection and validation to identify and mitigate noise-induced spurious drifts. The method’s performance will be evaluated against state-of-the-art drift detection methods across benchmark datasets. We will argue that the proposed method achieves a statistically significantly lower number of false positive drift detections over the existing approaches and better classifier accuracy in the presence of Gaussian noise.

Dr. Luca Olivieri

Ca' Foscari University, Italy



Pitfalls and Challenges of Blockchain Interoperability in a Fragmented Landscape

Blockchain technology, a decentralised and distributed ledger system, has revolutionised industries and digital cash systems by enabling transparent and tamper-proof transactions without the need for intermediaries. Interoperability has emerged as a critical concern in the evolution of the blockchain, particularly as the ecosystem becomes increasingly fragmented with the proliferation of diverse blockchain networks. The capacity of these distributed networks to exchange data and collaborate fluidly is fundamental to unlocking the broader capabilities of blockchain in domains such as decentralised finance (DeFi), supply chain logistics, digital identity, and beyond. Nevertheless, realising true interoperability presents numerous challenges and potential pitfalls. The keynote reflects on these ideas and gives an outlook on current developments, shortcomings and security issues.

Dr. Koushik Sinha

Southern Illinois University Carbondale, USA



Space, the Final Frontier: Unlocking the Potential of Orbital Networking and Edge Computing

The recent trend of deploying large-scale Low Earth Orbit (LEO) satellite constellations is catalyzing the development of a robust and efficient space-based data backhaul network. This promises a new era for networking and edge computing, where such orbital network and computing infrastructures promise significant reductions in end-to-end latency and ground station transmission loads. However, the unique characteristics of the space environment—including highly dynamic network topologies due to rapid satellite movement, harsh environment of space, varying atmospheric conditions—when combined with the inherent broadcast nature of wireless communication, introduce complexities absent in terrestrial networks. These challenges open up new and exciting research opportunities across all layers of the communication protocol stack for designing efficient, secure and reliable communication protocols while satisfying Quality of Service (QoS) constraints. Furthermore, realizing a practical satellite edge computing (SEC) framework on this dynamic network demands novel and innovative algorithmic solutions for crucial functions such as task management (scheduling, migration), fault tolerance (checkpointing, recovery), and coordination (consensus, group management), all while navigating the dynamic topology, and computational and storage constraints of the satellites. This talk will explore the current research landscape, highlight key opportunities, and discuss critical challenges within the rapidly evolving domains of space-based networking and satellite edge computing.

Dr. Utpal Garain

Indian Statistical Institute (ISI), Kolkata, India



Domain Learning, Explainability, and Scaling in GenAI Models

GenAI models achieved remarkable success across diverse domains, but fundamental questions remain about how models learn, explain, and scale. In this talk, we explore four interconnected aspects of deep learning. First, we examine domain-aware learning, demonstrating how vision models trained on 200,000 chess boards can infer patterns beyond object classification, hinting at implicit rule learning. Second, we discuss explainability in medical imaging, where end-to-end learning of expert-defined attributes improves interpretability and trustworthiness in AI-assisted diagnostics. Third, we address model reliability through expected calibration error (ECE), a measure of how well a model's confidence aligns with its correctness. Finally, we challenge conventional wisdom on scaling laws, presenting empirical evidence that, for specific tasks, an exponential scaling trend better describes performance improvements than the commonly assumed power-law relationship. Through these perspectives, we provide insights into the principles that govern GenAI models and their real-world implications.