Eventos pasados: Jean-Remi King. Towards a computational model of language processing in the adult human brain

What: Towards a computational model of language processing in the adult human brain

Where:  Zoom room 2

Who: Jean-Remi King, PhD (CNRS Researchers, Ecole Normale Supérieure, Facebook AI)

When:  Thursday, September 23rd at 1:30 PM.

Deep learning has recently allowed major progress in complex tasks such as summarization and text completion. Can these algorithms help us understand how the human brain processes language? Here, we investigate deep networks trained on a variety of tasks, ranging from character and speech recognition, to word and sentence embedding. We then test whether these networks represent text and speech similarly to the human brain, as recorded with magneto-encephalography (MEG, n=204), functional Magnetic Resonance Imaging (fMRI, n=589), and intracranial electrodes (n=176 patients, 20K electrodes). Our results show a systematic mapping between deep nets and the human brain: not only do these two systems share similar dynamics and representations, but the better the deep nets are, the more they resemble the brain. We show that this functional convergence provides a versatile framework to dissect language processing in the adult brain, from its hierarchical organization, to its syntactic and semantic modularization. More generally, these results illustrate how deep learning enables a radical shift in the neuroscientific enterprise and opens new avenues to understand the principles of intelligence.