Predictive Coding or Just Feature Discovery? An Alternative Account of Why Language Models Fit Brain Data

Shared functional specialization in transformer-based language models and the human brain

When processing language, the brain is thought to deploy specialized computations to construct meaning from complex linguistic structures. Recently, artificial neural networks based on the Transformer architecture have revolutionized the field of natural language processing. Transformers integrate contextual information across words via structured circuit computations. Prior work has focused on the internal representations ("embeddings") generated by these circuits. In this paper, we instead analyze the circuit computations directly: we deconstruct these computations into the functionally-specialized "transformations" that integrate contextual information across words. Using functional MRI data acquired while participants listened to naturalistic stories, we first verify that the transformations account for considerable variance in brain activity across the cortical language network. We then demonstrate that the emergent computations performed by individual, functionally-specialized "attention heads" differentially predict brain activity in specific cortical regions. These heads fall along gradients corresponding to different layers and context lengths in a low-dimensional cortical space.

Artificial Neural Network Language Models Predict Human Brain Responses to Language Even After a Developmentally Realistic Amount of Training

Artificial neural networks have emerged as computationally plausible models of human language processing. A major criticism of these models is that the amount of training data they receive far exceeds that of humans during language learning. Here, we use two complementary approaches to ask how the models' ability to capture human fMRI responses to sentences is affected by the amount of training data. First, we evaluate GPT-2 models trained on 1 million, 10 million, 100 million, or 1 billion words against an fMRI benchmark. We consider the 100-million-word model to be developmentally plausible in terms of the amount of training data given that this amount is similar to what children are estimated to be exposed to during the first 10 years of life. Second, we test the performance of a GPT-2 model trained on a 9-billion-token dataset to reach state-of-the-art next-word prediction performance on the human benchmark at different stages during training. Across both approaches, we find that (i) the models trained on a developmentally plausible amount of data already achieve near-maximal performance in capturing fMRI responses to sentences. Further, (ii) lower perplexity-a measure of next-word prediction performance-is associated with stronger alignment with human data, suggesting that models that have received enough training to achieve sufficiently high next-word prediction performance also acquire representations of sentences that are predictive of human fMRI responses. In tandem, these findings establish that although some training is necessary for the models' predictive ability, a developmentally realistic amount of training (∼100 million words) may suffice.

Lexical-Semantic Content, Not Syntactic Structure, Is the Main Contributor to ANN-Brain Similarity of fMRI Responses in the Language Network

Representations from artificial neural network (ANN) language models have been shown to predict human brain activity in the language network. To understand what aspects of linguistic stimuli contribute to ANN-to-brain similarity, we used an fMRI data set of responses to n = 627 naturalistic English sentences (Pereira et al., 2018) and systematically manipulated the stimuli for which ANN representations were extracted. In particular, we (i) perturbed sentences' word order, (ii) removed different subsets of words, or (iii) replaced sentences with other sentences of varying semantic similarity. We found that the lexical-semantic content of the sentence (largely carried by content words) rather than the sentence's syntactic form (conveyed via word order or function words) is primarily responsible for the ANN-to-brain similarity. In follow-up analyses, we found that perturbation manipulations that adversely affect brain predictivity also lead to more divergent representations in the ANN's embedding space and decrease the ANN's ability to predict upcoming tokens in those stimuli. Further, results are robust as to whether the mapping model is trained on intact or perturbed stimuli and whether the ANN sentence representations are conditioned on the same linguistic context that humans saw. The critical result-that lexical-semantic content 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.

Computational Language Modeling and the Promise of In Silico Experimentation

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.

Augmenting interpretable models with large language models during training

Recent large language models (LLMs), such as ChatGPT, have demonstrated remarkable prediction performance for a growing array of tasks. However, their proliferation into high-stakes domains and compute-limited settings has created a burgeoning need for interpretability and efficiency. We address this need by proposing Aug-imodels, a framework for leveraging the knowledge learned by LLMs to build extremely efficient and interpretable prediction models. Aug-imodels use LLMs during fitting but not during inference, allowing complete transparency and often a speed/memory improvement of greater than 1000x for inference compared to LLMs. We explore two instantiations of Aug-imodels in natural-language processing: Aug-Linear, which augments a linear model with decoupled embeddings from an LLM and Aug-Tree, which augments a decision tree with LLM feature expansions. Across a variety of text-classification datasets, both outperform their non-augmented, interpretable counterparts. Aug-Linear can even outperform much larger models, e.g. a 6-billion parameter GPT-J model, despite having 10,000x fewer parameters and being fully transparent. We further explore Aug-imodels in a natural-language fMRI study, where they generate interesting interpretations from scientific data.

Driving and suppressing the human language network using large language models

Transformer models such as GPT generate human-like language and are highly predictive of human brain responses to language. Here, using fMRI-measured brain responses to 1,000 diverse sentences, we first show that a GPT-based encoding model can predict the magnitude of brain response associated with each sentence. Then, we use the model to identify new sentences that are predicted to drive or suppress responses in the human language network. We show that these model-selected novel sentences indeed strongly drive and suppress activity of human language areas in new individuals. A systematic analysis of the model-selected sentences reveals that surprisal and well-formedness of linguistic input are key determinants of response strength in the language network. These results establish the ability of neural network models to not only mimic human language but also noninvasively control neural activity in higher-level cortical areas, like the language network.

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

Representations from artificial neural network (ANN) language models have been shown to predict human brain activity in the language network. To understand what aspects of linguistic stimuli contribute to ANN-to-brain similarity, we used an fMRI dataset of responses to n=627 naturalistic English sentences (Pereira et al., 2018) and systematically manipulated the stimuli for which ANN representations were extracted. In particular, we i) perturbed sentences' word order, ii) removed different subsets of words, or iii) replaced sentences with other sentences of varying semantic similarity. We found that the lexical semantic content of the sentence (largely carried by content words) rather than the sentence's syntactic form (conveyed via word order or function words) is primarily responsible for the ANN-to-brain similarity. In follow-up analyses, we found that perturbation manipulations that adversely affect brain predictivity also lead to more divergent representations in the ANN's embedding space and decrease the ANN's ability to predict upcoming tokens in those stimuli. Further, results are robust to whether the mapping model is trained on intact or perturbed stimuli, and whether the ANN sentence representations are conditioned on the same linguistic context that humans saw. The critical result-that lexical-semantic content 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.

Scaling laws for language encoding models in fMRI

Representations from transformer-based unidirectional language models are known to be effective at predicting brain responses to natural language. However, most studies comparing language models to brains have used GPT-2 or similarly sized language models. Here we tested whether larger open-source models such as those from the OPT and LLaMA families are better at predicting brain responses recorded using fMRI. Mirroring scaling results from other contexts, we found that brain prediction performance scales logarithmically with model size from 125M to 30B parameter models, with ~15% increased encoding performance as measured by correlation with a held-out test set across 3 subjects. Similar logarithmic behavior was observed when scaling the size of the fMRI training set. We also characterized scaling for acoustic encoding models that use HuBERT, WavLM, and Whisper, and we found comparable improvements with model size. A noise ceiling analysis of these large, high-performance encoding models showed that performance is nearing the theoretical maximum for brain areas such as the precuneus and higher auditory cortex. These results suggest that increasing scale in both models and data will yield incredibly effective models of language processing in the brain, enabling better scientific understanding as well as applications such as decoding.