Computational Language Modeling and the Promise of In Silico Experimentation

Abstract

Language neuroscience currently relies on two major experimental paradigms: controlled experiments using carefully hand-designed stimuli, and natural stimulus experiments. These approaches have complementary advantages which allow them to address distinct aspects of the neurobiology of language, but each approach also comes with drawbacks. Here we discuss a third paradigm—in silico experimentation using deep learning-based encoding models—that has been enabled by recent advances in cognitive computational neuroscience. This paradigm promises to combine the interpretability of controlled experiments with the generalizability and broad scope of natural stimulus experiments. We show four examples of simulating language neuroscience experiments in silico and then discuss both the advantages and caveats of this approach.

Driving and suppressing the human language network using large language models

“…In particular, an accurate model-to-brain encoding model can serve as a quantitative, assumption-neutral tool for deriving experimental materials aimed at understanding the functional organization of the language network and putatively downstream areas. Moreover, accurate encoding models can be used as a 'virtual language network' to simulate experimental contrasts in silico (e.g., 159,139,160 ). For prospective application, stimuli can be optimized for eliciting a strong response, thus allowing for efficient identification of language circuits, which may be especially important for individuals with brain disorders and other special populations, or in circumstances where time is of essence (e.g., neurosurgical planning and intraoperative testing).…”

“…In particular, an accurate model-to-brain encoding model can serve as a quantitative, assumption-neutral tool for deriving experimental materials aimed at understanding the functional organization of the language network and putatively downstream areas that support abstract knowledge and reasoning (e.g., 168,169,73,170 ). Moreover, accurate encoding models can be used as a 'virtual language network' to simulate experimental contrasts in silico (e.g., [171][172][173] ). In particular, the model-selected sentences can be queried in a highthroughput manner to analyze the response properties of the language network in detail, providing the ability to rapidly generate novel hypotheses about language processing that can then be tested in a 'closed-loop' manner.…”

Predicting the next sentence (not word) in large language models: What model-brain alignment tells us about discourse comprehension

“…Notably, we found higher model-brain alignment in the right IFG and IFGorb in the naturalistic reading of coherent sentences, while Jacoby and Fedorenko ( 24 ) found greater activation in these two regions for their unconnected sentences condition. This discrepancy points to an important difference between the contrast-based experimental fMRI approach and the model-based approach in our study ( 61 , 62 ). The former approach relies on contrasting and analyzing well-controlled conditions, so it may overlook the involvement of brain regions if the regions are active but do not show statistically stronger levels of activation compared to the contrasted condition.…”

Studying and improving reasoning in humans and machines

“…From the point of cognitive psychology, following successful antecedents in language processing and visual perception 26,[31][32][33][34] , it has been recently argued that LLMs may become useful tools to understand decision-making, learning and preferences [35][36][37] , to the point that it has been (possibly provokingly) claimed that LLMs experiments may at some point complement, if not replace, human psychological experiments 38,39 . Crucially, both these lines of reasoning (i.e., understanding LLMs cognitive abilities and validating LLMs are models of human reasoning) necessitate parallel investigation of human and machine performance with the same tools.…”