Abstract
Presentation Description :
Visual recognition is a fundamental function of the human brain, relying on a cascade of neural processes to transform low level inputs into semantic content. Despite significant advances in characterizing the locus and function of key perceptual cortical regions, integrating the temporal and spatial dynamics of this processing stream has been a challenge. In this talk, I will present a series of works which address this challenge by showing how the combination of MEG (or EEG), functional MRI measurements, representational geometry and deep neural networks can give new insights into visual processes in the human brain. First, I will present a novel method to characterize the interplay of feedforward and feedback mechanisms along the human ventral visual stream, and suggest how recurrent artificial neural networks can better explain the neural data in challenging visual tasks. Second, I will show how a deep generative adversarial autoencoder reveals dynamics of neural feedback processes as reconstructing low level visual information from high level latent representations. Finally, I will present the results of our latest study showing how recent developments in Large language models are the best to explain cross-modal perceptual similarities across vision and language.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.