"Deep language models as a cognitive model for natural language processing in the human brain"
Dr. Uri Hasson, Princeton Neuroscience Institute, Princeton University
Monday, February 6th at 5:30 p.m., Zoom
About Dr. Hasson
Uri Hasson received his Ph.D. in Neurobiology from the Weizmann Institute in Israel.
He was a postdoctoral fellow at NYU before moving to Princeton. His research program aimed to understand
how the brain processes real-life information and interacts with the environment, focusing on human face-to-face communication and natural language processing.
In recent years the research has been expanding to explore deep learning as a computational framework for
modeling the neural basis of cognition in real-life contexts and children's language development.
Abstract
Naturalistic experimental paradigms in cognitive neuroscience arose from a pressure to test, in real-world contexts,
the validity of models we derive from highly controlled laboratory experiments.
In many cases, however, such efforts led to the realization that models (i.e., explanatory principles) developed under particular experimental
manipulations fail to capture many aspects of reality (variance) in the real world. Recent advances in artificial neural networks provide an alternative
computational framework for modeling cognition in natural contexts. In contrast to the simplified and interpretable hypotheses we test in the lab,
these models learn how to act in the world from massive amounts of real-world examples (big data) by optimizing big models with millions to billions of parameters.
Surprisingly, such models' performance matches human performances on many cognitive tasks, including visual perceptions, language processing, and motor control.
At the same time, these models sacrifice understanding in favor of competence by being able to act without knowing why their choices are optimal or preferable in a given context.
In this talk, I will ask whether the human brain's underlying computations are similar or different from the underlying computations in deep neural networks.
The ability to think and reason using natural language separates us from other animals and machines.
In the talk, I will focus on the underlying neural process that supports natural language processing.
Our study aims to model natural language processing in the wild. To that end, we recorded intracranial EEG data from epileptic patients
while they engaged in open-ended free conversations with their doctors, friends, and family members during a week-long stay in the hospital.
I will provide evidence that our neural code shares some computational principles with deep language models. This indicates that, to some extent,
the brain relies on overparameterized optimization methods to comprehend and produce language. At the same time, I will present evidence that the brain
differs from deep language models as speakers try to convey new ideas and thoughts. Together, our findings expose some unexpected similarities to deep neural
networks while pointing to crucial human-centric missing properties in these machines.
Suggested Readings
Hasson, et al. (2020). Direct Fit to Nature: An Evolutionary Perspective on Biological and Artificial Neural Networks, Neuron, 105(3), 416-434. [.pdf]
Goldstein, et al. (2022). Shared computational principles for language processing in humans and deep language models, Nature Neuroscience 25, 369-380. [.pdf]