Giovani in un'ora - Ciclo di seminari - Quinta parte

Day - Time: 27 November 2024, h.11:00
Place: Area della Ricerca CNR di Pisa - Room: C-29
Speakers
Referent

Fabio Carrara

Abstract
Giulio Del Corso - "Generating a physically accurate cardiac MRI: a story of (interesting) failures and (justifiable) numerical shortcuts"

Abstract: "The generation of physically accurate cardiac-MRI video sequences is a challenging topic that requires a combination of modern generative approaches, classical methods of numerical analysis and signal processing techniques to achieve satisfactory results.

Although the subject is of great scientific interest, with applications ranging from the creation of synthetic databases for segmentation and classification to the generation of tools for training medical professionals, the scarcity of data available, individual variability and, above all, the presence of chaotic phenomena in cardiac fluid dynamics leave open opportunities for new approaches to the problem.

This presentation will therefore tell a story of failures, shortcuts and more or less legitimate tricks by integrating the power of modern deep learning approaches with the reliability of classical numerical analysis modelling in an attempt to obtain the first physically accurate 4D cardiac MRI."

Antonino Crivello - "Open Science Practices in Indoor Navigation Research"

Abstract: "The community of indoor positioning research has identified the need for a paradigm shift towards more reproducible and open research dissemination. Despite recent efforts to openly share data and code, accompanying research results with Open Research Data (ORD) is far from being the de facto standard option for publications in the indoor positioning field. The lack of recognized public benchmarks and the rather slow adoption of ORD, set a great volume of astute contributions in the field to remain irreproducible. In this talk, we review the landscape of Open Research Data in Indoor Positioning, enlisting, presenting, and analyzing the characteristic features of the relevant available open datasets of the field. We will also discuss potential shortcomings, and share lessons learned and observed good practices regarding the provision of a new ORD and the reuse of an existing one."

Giovanni Puccetti - "Outlier Dimensions in Large Language Models"

Abstract: "Transformer Based Language Models (LMs) are known to have outliers, specific dimensions in their hidden representations showing high magnitude when compared to others. These dimensions can compromise model performance when removed and they encode sufficient information to solve specific downstream tasks, such as language inference. This phenomenon is known for Encoder LMs, but it is less studied for Generative Large Language Models (LLMs), where longer training impacts the properties of outliers. This ongoing work aims to investigate the phenomenon of outliers in LLMs to understand which behaviours are preserved and which change. In particular we attempt to relate outliers to in context learning, one of the principal innovation provided by Generative LLMs."