Keynotes
Prof. Torsten Braun
Institute of Computer Science, University of Bern, Switzerland
Distributed Machine Learning for Future Mobile Communication Networks and Emerging Applications
Abstract: Distributed Machine Learning (DML) distributes model training and inference across different nodes to handle large data sets, to reduce training / inference time, and to limit raw data exchange, which also supports data privacy. In this keynote we will discuss several DML approaches such as federated learning, decentralized learning, split learning, and multi-agent reinforcement learning. We will show how those mechanisms can be used to address various problems in communications and networking such as channel estimation and beam-forming using massive Multiple Input Multiple Output (mMIMO) antenna systems and efficient placement of virtual network functions including orchestration of Radio Access Network (RAN) controllers. Moreover, we are introducing Mobile Mixed Reality Networks, where users and their devices are interacting and collaborating in mixed reality environments, such as for cooperative caching and sharing AI services such as viewport prediction, in order to achieve low delays required by mixed reality applications.
Prof. Hyunjoon Moon
Department of Computer Science and Engineering,
Sejong University, South Korea
Vision–Language Models: Architectures, Parameter-Efficient Adaptation, and Future Horizons
Abstract: Vision–Language Models (VLMs) have redefined multimodal artificial intelligence, driving major breakthroughs across image captioning, visual question answering, and cross-modal retrieval. Drawing from a systematic analysis of 115 critical papers published between 2018 and 2025, this keynote delivers a comprehensive roadmap of core VLM components, including pre-trained architectures, fine-tuning strategies, prompt engineering, adapter modules, and benchmarking datasets. We establish clear taxonomies and compare benchmark performance, with a dedicated focus on how lightweight, parameter-efficient adaptation methods minimize computational overhead in real-world deployment. Furthermore, the presentation critically examines architectural trade-offs, dataset-specific tuning, and prompt-based learning. Concluding with key challenges in scalability, generalization, and algorithmic bias, we explore emerging research frontiers including symbolic reasoning, multilingual adaptation, and energy-efficient VLM design. This talk serves as an essential framework for researchers and practitioners optimizing next-generation VLMs.