Giampaolo, Fabio (2023) Learning paradigms for neural networks: from backpropagation to locally backpropagated Forward-Forward. [Tesi di dottorato]

Anteprima

Testo Giampaolo_Fabio_36.pdf Download (8MB) | Anteprima

Abstract

This thesis delves into the transformative world of neural networks, integral to modern artificial intelligence, exploring their challenges and developments. It provides a comprehensive analysis of their fundamental components and learning processes, emphasizing the importance of a deep understanding of these elements in enhancing their practical utility. The work covers the intricacies of neural network learning, dissecting these systems to understand the dynamics of their training and operation. This insight is pivotal for adapting neural networks to solve challenging tasks and drive advancements in various fields. A major focus is on developing a robust forecasting framework for time series prediction, demonstrating the practical application and feasibility of neural learning. Additionally, the thesis explores alternative learning paradigms beyond traditional methods, highlighting new strategies that modify foundational training mechanisms. This includes an investigation into an alternative to backpropagation, showing its relevance and promise, especially in addressing federated learning challenges. In conclusion, the thesis underscores the necessity of understanding and innovating within the neural network domain. As technology evolves, this knowledge becomes critical in redefining what is achievable, with potential significant impacts across various scientific and technological fields.

Downloads

Actions (login required)

Modifica documento