Driving and suppressing the human language network using large language models

January 2024 Excited to share that our work on driving and suppressing brain responses in the language network is published in Nature Human Behavior (work with Aalok Sathe, Shashank Srikant, Maya Taliaferro, Mingye Wang, Martin Schrimpf, Kendrick Kay, and Ev Fedorenko).

In this work, we asked two main questions: i) Can we leverage GPT language models to drive or suppress brain responses in the human language network? ii) What is the “preferred” stimulus of the language network, and why?
To answer these questions, we recorded brain data while participants read 1,000 linguistically diverse sentences using fMRI. We fit an encoding model to predict the left hemisphere language network’s response to an arbitrary sentence from GPT embeddings. Next, we used our encoding model to identify new sentences to activate the language network maximally ("drive sentences") or minimally ("suppress sentences") by searching across a massive number of sentences (~1.8 million). We showed that these model-selected new sentences indeed drive and suppress the activity of human language areas in new individuals. Hence, we had trained an encoding model to generate predictions about the magnitude of activation in the language network for the new drive/suppress sentences and then closed the loop, so to speak, by collecting brain data for the new sentences from new participants, effectively “controlling” their brain activity.
Second, we wanted to understand what kinds of sentences that the language network responds to the most. This general approach is rooted in the work by Hubel and Wiesel in the late 1950’s and 1960’s, who discovered what kinds of visual input (e.g., specific orientations of light) that would make neurons in the visual cortex of cats and monkeys fire the most. In our study, we characterized all of our sentences (spanning the full range of brain responses from low to high) using 11 linguistic properties. We discovered that sentences with a certain degree of unusual grammar and/or meaning elicit the highest activity in the language network in the form of an inverted U-shape – for instance, sentences like “I’m progressive and you fall right.” would elicit the highest responses. Conversely, sentences that were highly normal/frequent (e.g., "We were sitting on the couch.") or highly unusual, like a string of words (e.g., "LG'll obj you back suggested Git.") would elicit quite low responses. In other words, the language network responds strongly to sentences that are “normal” (language-like) enough to engage it, but unusual enough to tax it. This means that specific areas of your brain – the language network – will respond selectively to linguistic input that aligns with the statistics of the language and will work really hard to make sense of the words and how they go together.

The paper can be found here: Tuckute, G., Sathe, A., Srikant, S., Taliaferro, M., Wang, M., Schrimpf, M., Kay, K., Fedorenko, E. (2024): Driving and suppressing the human language network using large language models, Nature Human Behavior, doi: https://doi.org/10.1038/s41562-023-01783-7.
I also wrote a blog post about this work for Springer Nature's "Behind the paper" series.

Many but not all deep neural network audio models capture brain responses and exhibit correspondence between model stages and brain regions

December 2023 Our work on models of the auditory cortex is hereby published in PLoS Biology (work with Jenelle Feather*, Dana Boebinger, and Josh McDermott). The paper presents an extensive analysis of audio neural networks as models of auditory processing in the brain.

We evaluated brain-model correspondence for 19 auditory deep neural networks (DNNs) (9 publicly available models, 10 models trained by us spanning four tasks) on two fMRI datasets (n=8, n=20) and using two different evaluation metrics (via regression and representational similarity analysis, RSA). We found that most DNNs (but not all) showed greater brain-model similarity than a traditional filter-bank model. Results were highly consistent between datasets and evaluation metrics. Most trained DNNs exhibited correspondence with the hierarchical organization of the auditory cortex, with earlier DNN stages best matching primary auditory cortex and later stages best matching non-primary cortex. Other two findings that are worth highlighting: i) A model’s training data significantly influences its similarity to brain responses: models trained to perform word/speaker recognition in the presence of background noise are much better at accounting for auditory cortex responses than models trained in quiet. ii) A model trained on multiple tasks was the best model of the auditory cortex overall, and also accounted for neural tuning to speech and music. Training on particular tasks improved predictions for specific types of tuning, with the MultiTask model getting “best of all worlds”.

Finally, our work demonstrates that if the core goal is to obtain the most quantitatively accurate model of the auditory system, machine learning models move us closer to this goal. However, our findings also highlight the explanatory gap that remains (predictions of all models are well below the noise ceiling), as well as the need for better model-brain evaluation metrics and finer-grained neural recordings to better distinguish models.

The paper can be found here: Tuckute, G.*, Feather*, J., Boebinger, D., McDermott, J. (2023): Many but not all deep neural network audio models capture brain responses and exhibit correspondence between model stages and brain regions, PLoS Biology 21(12), doi: https://doi.org/10.1371/journal.pbio.3002366.

Investigating brain-model similarity using information-theoretic compression

December 2023 Mycal Tucker and I will be presenting a poster at NeurIPS 2023 (UniReps Workshop; Dec 15, 3-5pm CT) on leveraging information theoretic tools to investigate similarity between brains and large language models (LLMs).

The motivation for our work stems from the observation that languages are efficient compressors -- we convey information via lossy representations (words) (e.g., Zaslavsky et al., 2018). Hence, a good model of human language processing should compress linguistic representations in ways similar to the human brain. LLMs are today’s most accurate models of human language processing, however it is unknown whether LLM representations contain similar amounts of information as the human brain. We took an initial stab at this question, and asked i) Do brain and LLM representations contain similar amounts of information?, and ii) Can we leverage an information bottleneck approach to generate compressed representations of brain activity and better unify the representational spaces of humans and LLMs during language processing?
Our prelimary results show that compressing brain activity in frontal language regions improves alignment with LLMs, suggesting that frontal regions in the human brain encode information that LLMs do not. There was no benefit to compressing temporal brain regions, suggesting that these regions might align better with the information present in current LLMs. Broadly, our work establishes the use of information theoretic tools as a "dial" to modify representational complexity and to better unify representational spaces, including those from biological and artificial lanuage processing.

The extended abstract can be found here: Tucker*, M. & Tuckute*, G., Increasing Brain-LLM Alignment via Information-Theoretic Compression, 37th Conference on Neural Information Processing Systems (NeurIPS 2023), UniReps Workshop.

Young Scientist Award at VCCA2023

June 2023 Very honored and thrilled to have received the Young Scientist Award at the Computational Audiology (VCCA) conference 2023! Thanks so much to the Computational Audiology community & conference organizers! I gave a talk on "Driving and suppressing the human language network using large language models" and took part in a panel discussion on large language models and chatbots. Highlights from the jury were: “Great clarity of explanations for very complex research. Excellent flow of presentation and live speaker (!). Scientific contribution, novelty and interdisciplinarity were outstanding and gave a view into the future. Great contributions to special session panel discussion, too.”. Thanks again!

Lexical semantic content, not syntactic structure, is the main contributor to ANN-brain similarity of fMRI responses in the language network

May 2023 Many studies have now established that large language models (LLMs) are predictive of human brain responses during language processing. Excited about our work that asks why and which aspects of the linguistic stimulus that contribute to LLM-brain similarity. Co-led with Carina Kauf*, and in collaboration with Roger Levy, Jacob Andreas, and Ev Fedorenko.

To answer these questions, we used an fMRI dataset of brain responses to n=627 diverse English sentences (Pereira et al., 2018) and systematically manipulated the stimuli for which LLM representations were extracted. We then fitted standard encoding model to predict brain activity from the LLM representations of these perturbed stimuli, and evaluated the resulting predictivity performance.
Critically, we found that the lexical-semantic content of the sentence rather than the sentence’s syntactic form is primarily responsible for the LLM-to-brain similarity. This means that manipulations that remove the content words or alter the meaning of the sentence decrease brain predictivity. Conversely, manipulations that remove function words or perturb the syntactic structure of the sentence (e.g., by shuffling the word order) do not lead to large decreases in brain predictivity. We show that stimulus manipulations that adversely affect brain predictivity have two interpretable causes: i) they lead to more divergent representations in the LLM’s embedding space (relative to the representations of the original stimuli), and decrease the LLM’s ability to predict upcoming tokens in those stimuli. So, stimuli that are less predictable on average lead to larger decreases in brain predictivity performance.
In summary, the critical result—that lexical-semanticcontent is the main contributor to the similarity between ANN representations and neural ones—aligns with the idea that the goal of the human language system is to extract meaning from linguistic strings. Finally, this work highlights the strength of systematic experimental manipulations for evaluating how close we are to accurate and generalizable models of the human language network.

The preprint can be found here: Kauf*, C, Tuckute*, G., Levy, R., Andreas, J., Fedorenko, E. (2023): Lexical semantic content, not syntactic structure, is the main contributor to ANN-brain similarity of fMRI responses in the language network, bioRxiv2023.05.05.539646; doi: https://www.biorxiv.org/content/10.1101/2023.05.05.539646v1. The paper will appear in the "Cognitive computational neuroscience of language" special issue of Neurobiology of Language.

Driving and suppressing the human language network using large language models

April 2023 Very excited to share our work on leveraging the predictive power of large language models to drive and suppress brain responses in the human language network. In collaboration with Aalok Sathe, Shashank Srikant, Maya Taliaferro, Mingye (Christina) Wang, Martin Schrimpf, Kendrick Kay, and Ev Fedorenko.

Transformer language models are today's most accurate models of language processing in the brain (e.g., Schrimpf et al., 2021; Goldstein et al., 2022; Caucheteux et al., 2023). However, there has been no attempt to test whether LLMs can causally control language responses in the brain. Recent work in visual neuroscience has shown that artificial neural network models of image recognition can causally intervene on the non-human primate visual system by generating visual stimuli that modulate activity in different regions of the ventral visual pathway. In this work, we ask whether similar model-based control is feasible for the higher-level cognitive domain of language: can we leverage the predictive power of LLMs to identify new stimuli to maximally drive or suppress brain responses in the language network of new individuals?

We developed an encoding model of the left hemisphere language network in the human brain with the goal of identifying new sentences that would activate the language network to a maximal or minimal extent. To do so, we recorded brain data while participants read 1,000 linguistically diverse sentences using fMRI. We fitted an encoding model to predict the language network’s response to an arbitrary sentence from GPT2-XL embeddings. Using our encoding model, we identified novel sentences to activate the language network maximally (drive sentences) or minimally (suppress sentences). We searched across ~1.8M sentences to identify these novel drive and suppress sentences and recorded brain data in new individuals using two different fMRI designs (event-related vs blocked). We found that the model-selected sentences indeed drive and suppress brain responses in new individuals as predicted, demonstrating non-invasive control of language responses in the brain.
Finally, why do certain sentences elicit higher responses than others? We took advantage of the broad distribution of model-selected linguistic input and obtained 11 features to characterize our experimental sentences: Surprisal and 10 behavioral norms (collected across=3,600 participants) reflecting 5 broad aspects of linguistic input (form/meaning, content, emotion, imageability, perceived frequency). We found that sentences that are surprising, fall in the middle of the grammaticality and plausibility range, and are hard to visualize elicit the strongest responses in the language network.
In summary, we demonstrate that it is possible to drive activity in the language network up or down with model (GPT2-XL)-selected stimuli in new individuals. Hence, we establish LLMs as causal tools to investigate language in the mind and brain, with implications for future basic-science investigations and clinical research.

The preprint can be found here: Tuckute, G., Sathe, A., Srikant, S., Taliaferro, M., Wang, M., Schrimpf, M., Kay, K., Fedorenko, E. (2023): Driving and suppressing the human language network using large language models, bioRxiv2023.04.16.537080; doi: https://doi.org/10.1101/2023.04.16.537080.

Cortical registration using anatomy and function

March 2023 We released a preprint on how to align brains using anatomy and function (to appear at MIDL 2023). This project is led by Jian (Andrew) Li, and in collaboration with Ev Fedorenko, Brian L. Edlow, Bruce Fischl*, and Adrian V. Dalca*.

Brains have complex geometric anatomy and brains differ a lot across individuals. Aligning a brain to another individual or to a common atlas space is a challenging task. Traditionally, this challenge has been solved using registration of anatomical folding patterns of the cortex. However, it is known that many functional regions in the brain do not exhibit a consistent mapping onto macro-anatomical landmarks. Hence, we might miss out on crucial information in aligning brains if we do not take function into account. In this work, we propose JOSA (Joint Spherical registration and AAtlas building that models anatomical and functional differences when aligning brains and building cortical atlases. As a proof of concept, we used functional data from a well-validated language localizer task (e.g., Lipkin et al., 2022) for a set of 150 participants, jointly aligning the language contrast (sentences versus strings of non-words) with cortical folding patterns, demonstrating better alignment in both folding patterns and function compared to two existing methods.
Upon publication, the method will be integrated as a part of VoxelMorph and/or FreeSurfer.

The preprint can be found here: Li, A., Tuckute, G., Fedorenko, E., Edlow, B.L., Fischl, B., Dalca, A.V. (2023): Joint cortical registration of geometry and function using semi-supervised learning, arXiv2303.01592; doi: https://doi.org/10.48550/arXiv.2303.01592.

Preprint on deep neural network models of the auditory system is out

September 2022 We released the preprint of our work on how deep neural networks (DNNs) for audio can account for brain responses in the human auditory cortex. This project is co-led with Jenelle Feather, and in collaboration with Dana Boebinger and Josh McDermott.

We evaluated brain-model correspondence for 19 DNNs (9 publicly available models, 10 models trained by us spanning four tasks) on two fMRI datasets (n=8, n=20) and using two different evaluation metrics (via regression and representational similarity analysis, RSA). We make the following four main claims: 1) Most DNN models (but not all!) outperformed traditional models of the auditory cortex. Results were highly consistent between datasets and evaluation metrics. The overall best DNN model was trained on multiple tasks (word, speaker, environmental sound recognition). 2) This brain-DNN similarity was strictly dependent on task-optimization. DNNs with permuted weights (which destroys the structure learned during model training) performed below the baseline model. 3) Most DNNs exhibited systematic correspondence with the hierarchical organization of the auditory cortex, with earlier DNN stages best matching primary auditory cortex and later stages best matching non-primary cortex. This was not true for permuted networks. 4) The task a DNN model is trained on influences its match to the brain, with e.g., speech-trained models best matching cortical speech responses (scatter plot on the right). 5) Finally, in light of recent discussion suggesting that the dimensionality of a model’s representation correlates with regression-based brain predictions, we evaluated how the effective dimensionality (ED) of each network stage correlated with both the regression and RSA metrics. There was a modest correlation between ED and brain-model similarity but significantly less than that between the two datasets or the two similarity measures. Thus ED does not seem to explain most of the variance across DNNs in our datasets.
Overall, we demonstrate that many, but not all, DNN models account for responses in the human auditory cortex with hierarchical stage-region correspondence, and provide some hints of how to improve brain-model matches for future models.

The preprint can be found here: Tuckute, G.*, Feather*, J., Boebinger, D., McDermott, J. (2022): Many but not all deep neural network audio models capture brain responses and exhibit hierarchical region correspondence, bioRxiv2022.09.06.506680; doi: https://doi.org/10.1101/2022.09.06.506680.