קולוקוויום מחקרי - החוג לבלשנות

במסגרת הקולוקוויום המחקרי של החוג לבלשנות, שמתקיים מדי יום חמישי, בין השעות 16:15-17:45 בבניין ווב (חדר 103):

הרצאתו של יאיר לקרץ מ-École normale supérieure - PNS, שכותרתה:

Deep Neural Models as a Window into a Neural Theory of Language: Successes and Challenges

תקציר:

Neuroimaging and electrophysiological studies have localized where in the brain language is processed and identified the typical processing times. However, the how, the computational mechanisms underlying language processing, remain largely unknown. One of the main challenges to achieve a neural theory of language processing is that linguistic theory is often described in terms of discrete symbolic representations, whereas neural theory is described in terms of continuous variables, such as neural activity or synaptic weights. Recent advances in deep neural models now provide a compelling approach to bridge this gap. Neural models are trained on extensive text corpora and demonstrate impressive language capabilities, on a variety of linguistic tasks. These models, like the human brain, are based on vectorial representations and provide full access to their neural activity and connectivity during language processing, which is infeasible to achieve in humans. Neural models therefore provide a new opportunity to achieve a neural theory of language. This talk will examine how these models might serve as a 'window' into neural language processing. We will discuss our studies that utilize these models to investigate the neural computations underlying language, and how their internal representations may offer new insights into how the human brain processes language. Finally, we will discuss remaining challenges of this approach, acknowledging both the promise and the limitations of deep neural models in contributing to a neural theory of language.

הציבור מוזמן!

קישור להקלטה

קישור לתכנית הקולוקוויום המלאה