Seminario
Everything you always wanted to know about ChatGPT but were afraid to as
Abstract dell'evento
The Transformer architecture, introduced in 2017 by Vaswani and other Google researchers, underpins Large Language Models (LLMs), which have reached a level of linguistic competence comparable to that of humans. These models also demonstrate remarkable abilities in logical reasoning, creative problem-solving, contextual understanding, and the synthesis of complex knowledge, reflecting multiple aspects of human cognitive skills. In this talk, I will focus specifically on decoder-only Transformers during the inference phase, that is, when they generate responses in normal operation. I will not consider their functioning during the training phase. While the Transformer’s structure remains the same in both inference and training, understanding the principles behind inference is particularly insightful and debunks the false notion that these models are purely statistical. Instead, I will argue for an interpretation of the model’s structure and functioning that better captures its essential aspects and provides a unified perspective on the diverse processes that constitute the system as a whole. This interpretation can be summarized as follows: a Transformer is a complex system of neural networks, interconnected in a non-standard way, that transforms concepts through concepts.
Programma
Il seminario fa parte del ciclo: Serpieri Lectures (Ciclo di Lezioni per i dottorandi ReMeST)
Relatori/Relatrici
MARCO GIUNTI (University of Cagliari)
Dettagli sull'evento
Data e luogo
Inizio: 26/03/2025
alle ore 16:00
Fine: 26/03/2025
alle ore 18:00
Collegio Raffaello (Urbino, Piazza della Repubblica, 13) Aula Magna
Organizzato e promosso da:
Dipartimento di Scienze Pure e Applicate
Scuola di Scienze, Tecnologie e Filosofia dell'Informazione
Modalità di partecipazione
Altre informazioni utili
Ingresso libero.
Per informazioni: alessandra.petronio@uniurb.it
Per ricevere il link zoom: giovanni.macchia@uniurb.it
