Na próxima 4ª feira, dia 29 de novembro, teremos duas talks no âmbito do processo de scouting.

Autor e vínculos: António Guilherme Correia, University of Jyväskylä, Faculty of Information Technology

Bio: António Correia is a Postdoctoral Researcher and member of the teaching staff in the Faculty of Information Technology at the University of Jyväskylä, Finland, where he conducts research on the intersectional space of human-artificial intelligence (AI) interaction in multiple research settings. Moreover, he is also a Research Assistant at the Institute for Systems and Computer Engineering, Technology and Science (INESC TEC), Porto, Portugal. He holds a Ph.D. in Computer Science and an M.Sc. in Information and Communication Technologies, with Summa Cum Laude honours, from the University of Trás-os-Montes and Alto Douro, Vila Real, Portugal. António was the first Portuguese to get awarded the prestigious Microsoft Research Ph.D. Fellowship. Besides his experience as a Researcher at Microsoft, he formerly worked as a Visiting Scholar at the University of Nebraska at Omaha, College of Information Science & Technology, NE, USA. Furthermore, he was also a Postgraduate Visiting Researcher at the University of Kent, UK. António holds more than ten years of experience in research, and his publication record comprises over 50 peer-reviewed papers in top-tier conferences and journals such as Artificial Intelligence Review, IEEE Transactions on Human-Machine Systems, User Modeling and User-Adapted Interaction, and Scientometrics. In line with this, he has also served as a scientific committee member for several venues covering aspects of computer science and has been consistently involved in a number of projects conducted at the international level, for instance, in collaboration with Carnegie Mellon University within the project "eCSAAP - expert Crowdsourcing for Semantic Annotation of Atmospheric Phenomena" (Ref.: CMU/CS/0012/2017).

Título: Contemporary Challenges in Designing Human-AI Interaction for Scientific Discovery

Abstract: Interdisciplinary research efforts in human-powered artificial intelligence (AI) algorithms have contributed to reveal hidden relationships and properties across large-scale knowledge graphs. However, the expanding scientific landscape presents challenges in capturing knowledge flows and their impact accurately when considering the complex network of interrelationships that characterize the scientific activity. For a long time, scholars have made progress in the quantitative analysis of science indicators, historical footprints, and network dynamics. Despite the remarkable strides over the last decade, the pipeline underlying the measurement of scientific output is still difficult to execute in a fully automated way. To overcome these challenges, several academics and practitioners have increasingly explored hybrid intelligent systems that combine human and machine intelligence in data-driven research as an instrument of science policy. Building upon these promising advancements, this talk presents a reinforcement learning from human feedback (RLHF) approach, offering insights for implementing hybrid socio-technical algorithmic systems intended to support research evaluation and strategic decision-making while enabling new forms of human-AI interactive and continuous sensemaking.

-

Autor e vínculos: Marcos Treviso, Instituto de Telecomunicações

Bio: Marcos Treviso completed his Ph.D. with Distinction and Honour at IST/University of Lisbon, advised by Prof. André Martins, focusing on the role of sparsity in interpretability in NLP, particularly in Machine Translation Quality Estimation. His work extended to Efficient Transformers and Continuous Attention Mechanisms, contributing to several papers for conferences like NeurIPS and ACL. Marcos also holds an M.Sc. in Computer Science and Computational Mathematics from the University of São Paulo (USP), where he researched neural networks in speech transcript analysis under Prof. Sandra M. Aluísio.

Título: Efficient, Transparent, and User-Centric Methods for Long-Context NLP
Abstract: Current natural language processing (NLP) models are hindered by their dependence on over-parameterized black boxes, raising concerns about their reliability, confidence, and fairness. While several explainability approaches have been proposed for shedding light into neural networks' decisions, ranging from built-in (e.g., attention mechanisms) to post-hoc methods (e.g., gradient-based measures), their evaluation often sidesteps the crucial aspect of effectively communicating the underlying model behavior to humans. In this presentation, I outline my research journey, focusing on neural network interpretability through simulability and sparsity. My PhD work centers on creating frameworks to evaluate and improve explainability methods, which contributed to notable successes in two editions of the Explainable Machine Translation Quality Estimation Shared Task. Advancing to future prospects, I highlight the importance of long-context NLP methods, emphasizing their growing relevance and challenges in tasks such as language modeling and document-level machine translation. Large Language Models (LLMs) like GPT-3 and GPT-4 are presented as particularly successful cases of long-context NLP. The discussion integrates themes of efficiency, interpretability, and user-centric design, aiming to inspire innovations in natural language processing and machine learning.