Bio-Inspired Methods for Deep Learning
-
Day - Time:
22 March 2024, h.15:30
-
Place:
Area della Ricerca CNR di Pisa - Room: Faedo (C-29)
Speakers
Referent
Giuseppe Amato
Abstract
Bio-Inspired Deep Learning (BIDL) represents a promising and currently active research area, at the intersection of computer science and neuroscience. While, on one hand, recent work aims at enhancing current deep learning technologies with more biologically inspired features, on the other hand, engineering insights can guide the search of biological structures and mechanisms from which complex cognitive abilities emerge. In this presentation, I will illustrate some of the limitations of current deep learning technologies, such as energy and data inefficiency, lack of robustness to certain types of noise, limited interpretability, and I will discuss a few directions of research, in the field of BIDL, that can address such limitations. Starting from synaptic plasticity models for improved learning abilities, I will move to spiking neural computation models, which enable more energy-efficient computation. Finally, I will discuss recent learning algorithms that represent interesting alternatives to traditional backpropagation-based training.