# 140 How the Brain Processes Language, With MIT Neuroscientist Ev Fedorenko

Joining SlatorPod this week is Ev Fedorenko, Associate Professor of Neuroscience at the Department of Brain and Cognitive Sciences, MIT. Ev also runs EvLab, an MIT language lab that discovers how the human brain creates language.

Ev talks about the different hypotheses concerning the origin of language and how it has likely been a gradual evolution. She shares a number of intriguing research findings on the relationship between language and abstract representations of structure (i.e., complex thought).

Ev discusses how language processing takes place and how we can use brain imaging to compare language with other non-linguistic tasks, such as solving math problems and composing music. She questions whether specific languages can be objectively easy or difficult to learn as an adult.

She also considers what sets polyglots apart when it comes to learning languages and some of the generalizations made in research. Ev talks about how language processing in machines like GPT-3 compares to that in humans. She argues that it would be more fruitful to build language systems that are structured similar to the human brain.

Ev concludes with the collaboration between academia and the booming field of applied AI, despite different goals. She touches on the MIT Quest for Intelligence, which brings together scientists and engineers to build better human-like models for the benefit of society.

Chapter Markers:
00:00:00 Intro and Agenda
00:01:01 Ev's Research
00:02:10 Ev's Background
00:03:20 The Origin of Language
00:07:42 The Relationship Between Language and Thought
00:11:46 What Regions of the Brain Support Language Processing?
00:14:41 Processing Word Meanings and Syntax
00:16:58 How Much "Compute" Goes Into Human Language Processing?
00:18:48 Easy-to-Learn vs. Difficult-to-Learn Languages
00:21:29 Mother Tongue vs. Acquired Fluency
00:25:13 The Phenomenon of Polyglots
00:29:30 The Process of Translation
00:32:38 Language Processing in Machines vs. Humans
00:38:30 Collaboration Between Academia and Industry

[BrianRandall-dypv]

Excellent overview and breaking down our cognitive ability around language. Awesome!

[thormay]

So LLMs have systematized their knowledge of word collocations in formats that enable them to predict or emulate the way humans use language. The human use of language, from the research of people like Ev Fedorenko, is cognitively distinct from human thinking processes. Human thinking processes exhibit knowledge of the world in a manner unavailable to the lexicalized syntax of LLMs. It now seems (to me) that the most challenging question for future research might be how LLM-like constructions could systematize "their knowledge of the world which they are given access to" in ways that might predict or emulate the ways humans think. ( Just a reflection from one of the walking wounded :). In the 1980s & 1990s I walked away from two PhD candidatures in Chomsky-like generative linguistics after intuiting that the models couldn't work in principle, without being clever enough to invent an alternative. Wrote one on something else).