Keynotes
Harnessing Large Language Models for Entity Processing in
Resource-Constrained Environments
 |
Toshiyuki Amagasa
|
Abstract
Data preprocessing, particularly named-entity recognition (NER) and entity matching (EM), has long posed significant challenges in data engineering. While traditional methods have made strides, the advent of large language models
(LLMs) has ushered in a new era of possibilities, potentially surpassing conventional approaches in both NER and EM tasks. However, the practical application of LLMs
faces several hurdles, most notably the constraints of limited training data and computational resources.
This talk explores these challenges and discusses strategies to leverage the power of LLMs for entity processing within resource-limited settings. We will delve into 1) the current landscape of NER and EM in data preprocessing, 2) the
transformative potential of LLMs in these areas and teal-world constraints hampering LLM adoption, and 3) promising approaches to optimize LLM performance under
resource limitations
Bio
Toshiyuki Amagasa received B.E., M.E., and Ph.D from the Department of Computer Science, Gunma University in 1994, 1996, and 1999, respectively. He is currently a full professor at the Center for Computational Sciences, University of Tsukuba. His research interests cover database systems, data mining, and database application in scientific domains. He is a senior member of IPSJ, IEICE, and IEEE, a board member of DBSJ, and a member of ACM.
Towards Explainable AI-Powered Malware Detection
 |
Martin HomolaComenius University Bratislava, Slovakia |
Abstract
Malware analysis has become more and more challenging due to the rapid evolution of attack techniques, and an ever increasing volume of new malware samples. Indeed traditional signature-based approaches struggle with this load and machine learning has steadily gained ground in this application area due to its effectiveness and classification power. However, much as in other domains, malware experts lack the confidence and trust in opaque machine learning models that provide classification of the sample but little or no justifications to support and explain such decisions. The lecture will delve into the novel explainable AI methods and techniques and illustrate their benefits and the trade-offs on the malware detection use case.
Bio
Martin Homola earned his PhD from Comenius University Bratislava, Faculty of Mathematics Physics and Informatics in 2010, in the area of distributed ontologies. During 2009–2012 he was a postdoc at Fundazione Bruno Kessler, Trento, Italy, where he worked in the Data & Knowledge Management Group on contextualized knowledge representation. After returning to his alma mater, he joined the Department of Applied Informatics. He earned his habilitation in 2018. He currently serves as the head of the KR research group, the head of the AI section of the department, and as the vice-dean for IT, PR and industrial collaboration of the faculty. His current research interests include knowledge representation in general, explainable AI, and applications of knowledge representation in information security, education, and other areas.
Automatic Speech Recognition in Adverse Acoustic Conditions
 |
Antonio Liotta
|
Abstract
Automatic speech recognition (ASR) is crucial to the development of effective voice-controlled applications. In fact, significant progress has been made in the last few years, and ASR multi-language models are readily available. But have these really reached near-human accuracy? And what is missing in the quest for super-human performance under broad acoustic conditions? In this talk, I revisit the status of ASR systems, which are typically trained on millions of hours of ordinary voice samples. I consider the strength of some of the most popular models, looking also at their promises and limitations. To get a more complete panorama of ASRs, I explore their effectiveness under a range of adverse acoustic conditions. As it turns out, ASRs do not yet meet some essential requirements, particularly when it comes to voices that are somewhat impaired or pathological. By means of a recent pilot study on “ASR for pathological voices”, I discuss experimental results, considering diverse models, languages and datasets. ASRs do degrade rapidly in face of typical disturbances, pointing to a number of interesting research propositions.
Bio
Antonio Liotta is Full Professor at the Faculty of Engineering, Free University of Bolzano (Italy), where he teaches Data Science and Machine Learning. He is the chairman of the Data-Driven Artificial Intelligence research area and the director of the PhD school in Computer Science.
Antonio’s passion for artificial intelligence, has driven his academic career through the meanders of artificial vision, e-health, intelligent networks and intelligent systems. Antonio’s team is renowned for his contributions to micro-edge intelligence and miniaturized machine learning, which have significant potential in harnessing data-intensive systems, for instance in the context of smart cities, cyber-physical systems, Internet of Things, smart energy, and machine learning with humans in the loop. He has led the international team that has recently made a breakthrough in artificial neural networks, initiating a new research strand on “sparse neural networks for embedded learning”. Antonio was the founding director of the Data Science Research Centre at the University of Derby. He has set up several cross-border virtual teams, and has been credited with over 350 publications involving, overall, more than 150 co-authors. Antonio is Editor-in-Chief of the Springer Internet of Things book series , and associate editor of several prestigious journals. He is co-author of the books “Networks for Pervasive Services: six ways to upgrade the Internet” and “Data Science and Internet of Things”.