The neural architecture of language: Integrative modeling converges on predictive processing

Dissociating language and thought in large language models

Large language models (LLMs) have come closest among all models to date to mastering human language, yet opinions about their linguistic and cognitive capabilities remain split. Here, we evaluate LLMs using a distinction between formal linguistic competence (knowledge of linguistic rules and patterns) and functional linguistic competence (understanding and using language in the world). We ground this distinction in human neuroscience, which has shown that formal and functional competence rely on different neural mechanisms. Although LLMs are surprisingly good at formal competence, their performance on functional competence tasks remains spotty and often requires specialized fine-tuning and/or coupling with external modules. We posit that models that use language in human-like ways would need to master both of these competence types, which, in turn, could require the emergence of separate mechanisms specialized for formal versus functional linguistic competence.

An inclusive multivariate approach to neural localization of language components

To determine how language is implemented in the brain, it is important to know which brain areas are primarily engaged in language processing and which are not. Existing protocols for localizing language are typically univariate, treating each small unit of brain volume as independent. One prominent example that focuses on the overall language network in functional magnetic resonance imaging (fMRI) uses a contrast between neural responses to sentences and sets of pseudowords (pronounceable nonwords). This contrast reliably activates peri-sylvian language areas but is less sensitive to extra-sylvian areas that are also known to support aspects of language such as word meanings (semantics). In this study, we assess areas where a multivariate, pattern-based approach shows high reproducibility across multiple measurements and participants, identifying these areas as multivariate regions of interest (mROI). We then perform a representational similarity analysis (RSA) of an fMRI dataset where participants made familiarity judgments on written words. We also compare those results to univariate regions of interest (uROI) taken from previous sentences > pseudowords contrasts. RSA with word stimuli defined in terms of their semantic distance showed greater correspondence with neural patterns in mROI than uROI. This was confirmed in two independent datasets, one involving single-word recognition, and the other focused on the meaning of noun-noun phrases by contrasting meaningful phrases > pseudowords. In all cases, areas of spatial overlap between mROI and uROI showed the greatest neural association. This suggests that ROIs defined in terms of multivariate reproducibility can help localize components of language such as semantics. The multivariate approach can also be extended to focus on other aspects of language such as phonology, and can be used along with the univariate approach for inclusively mapping language cortex.

Language is primarily a tool for communication rather than thought

Language is a defining characteristic of our species, but the function, or functions, that it serves has been debated for centuries. Here we bring recent evidence from neuroscience and allied disciplines to argue that in modern humans, language is a tool for communication, contrary to a prominent view that we use language for thinking. We begin by introducing the brain network that supports linguistic ability in humans. We then review evidence for a double dissociation between language and thought, and discuss several properties of language that suggest that it is optimized for communication. We conclude that although the emergence of language has unquestionably transformed human culture, language does not appear to be a prerequisite for complex thought, including symbolic thought. Instead, language is a powerful tool for the transmission of cultural knowledge; it plausibly co-evolved with our thinking and reasoning capacities, and only reflects, rather than gives rise to, the signature sophistication of human cognition.

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

Current large language models (LLMs) rely on word prediction as their backbone pretraining task. Although word prediction is an important mechanism underlying language processing, human language comprehension occurs at multiple levels, involving the integration of words and sentences to achieve a full understanding of discourse. This study models language comprehension by using the next sentence prediction (NSP) task to investigate mechanisms of discourse-level comprehension. We show that NSP pretraining enhanced a model's alignment with brain data especially in the right hemisphere and in the multiple demand network, highlighting the contributions of nonclassical language regions to high-level language understanding. Our results also suggest that NSP can enable the model to better capture human comprehension performance and to better encode contextual information. Our study demonstrates that the inclusion of diverse learning objectives in a model leads to more human-like representations, and investigating the neurocognitive plausibility of pretraining tasks in LLMs can shed light on outstanding questions in language neuroscience.

Predictive language processing: integrating comprehension and production, and what atypical populations can tell us

Predictive processing, a crucial aspect of human cognition, is also relevant for language comprehension. In everyday situations, we exploit various sources of information to anticipate and therefore facilitate processing of upcoming linguistic input. In the literature, there are a variety of models that aim at accounting for such ability. One group of models propose a strict relationship between prediction and language production mechanisms. In this review, we first introduce very briefly the concept of predictive processing during language comprehension. Secondly, we focus on models that attribute a prominent role to language production and sensorimotor processing in language prediction ("prediction-by-production" models). Contextually, we provide a summary of studies that investigated the role of speech production and auditory perception on language comprehension/prediction tasks in healthy, typical participants. Then, we provide an overview of the limited existing literature on specific atypical/clinical populations that may represent suitable testing ground for such models-i.e., populations with impaired speech production and auditory perception mechanisms. Ultimately, we suggest a more widely and in-depth testing of prediction-by-production accounts, and the involvement of atypical populations both for model testing and as targets for possible novel speech/language treatment approaches.

A unifying framework for functional organization in early and higher ventral visual cortex

A key feature of cortical systems is functional organization: the arrangement of functionally distinct neurons in characteristic spatial patterns. However, the principles underlying the emergence of functional organization in the cortex are poorly understood. Here, we develop the topographic deep artificial neural network (TDANN), the first model to predict several aspects of the functional organization of multiple cortical areas in the primate visual system. We analyze the factors driving the TDANN's success and find that it balances two objectives: learning a task-general sensory representation and maximizing the spatial smoothness of responses according to a metric that scales with cortical surface area. In turn, the representations learned by the TDANN are more brain-like than in spatially unconstrained models. Finally, we provide evidence that the TDANN's functional organization balances performance with between-area connection length. Our results offer a unified principle for understanding the functional organization of the primate ventral visual system.

Development of a Self-Report Measure of Prediction in Daily Life: The Prediction-Related Experiences Questionnaire

PURPOSE

Predictions are complex, multisensory, and dynamic processes involving real-time adjustments based on environmental inputs. Disruptions to prediction abilities have been proposed to underlie characteristics associated with autism. While there is substantial empirical literature related to prediction, the field lacks a self-assessment measure of prediction skills related to daily tasks. Such a measure would be useful to better understand the nature of day-to-day prediction-related activities and characterize these abilities in individuals who struggle with prediction.

METHODS

An interdisciplinary mixed-methods approach was utilized to develop and validate a self-report questionnaire of prediction skills for adults, the Prediction-Related Experiences Questionnaire (PRE-Q). Two rounds of online field testing were completed in samples of autistic and neurotypical (NT) adults. Qualitative feedback from a subset of these participants regarding question content and quality was integrated and Rasch modeling of the item responses was applied.

RESULTS

The final PRE-Q includes 19 items across 3 domains (Sensory, Motor, Social), with evidence supporting the validity of the measure's 4-point response categories, internal structure, and relationship to other outcome measures associated with prediction. Consistent with models of prediction challenges in autism, autistic participants indicated more prediction-related difficulties than the NT group.

CONCLUSIONS

This study provides evidence for the validity of a novel self-report questionnaire designed to measure the day-to-day prediction skills of autistic and non-autistic adults. Future research should focus on characterizing the relationship between the PRE-Q and lab-based measures of prediction, and understanding how the PRE-Q may be used to identify potential areas for clinical supports for individuals with prediction-related challenges.

Natural language instructions induce compositional generalization in networks of neurons

A fundamental human cognitive feat is to interpret linguistic instructions in order to perform novel tasks without explicit task experience. Yet, the neural computations that might be used to accomplish this remain poorly understood. We use advances in natural language processing to create a neural model of generalization based on linguistic instructions. Models are trained on a set of common psychophysical tasks, and receive instructions embedded by a pretrained language model. Our best models can perform a previously unseen task with an average performance of 83% correct based solely on linguistic instructions (that is, zero-shot learning). We found that language scaffolds sensorimotor representations such that activity for interrelated tasks shares a common geometry with the semantic representations of instructions, allowing language to cue the proper composition of practiced skills in unseen settings. We show how this model generates a linguistic description of a novel task it has identified using only motor feedback, which can subsequently guide a partner model to perform the task. Our models offer several experimentally testable predictions outlining how linguistic information must be represented to facilitate flexible and general cognition in the human brain.

The language network as a natural kind within the broader landscape of the human brain

Language behaviour is complex, but neuroscientific evidence disentangles it into distinct components supported by dedicated brain areas or networks. In this Review, we describe the 'core' language network, which includes left-hemisphere frontal and temporal areas, and show that it is strongly interconnected, independent of input and output modalities, causally important for language and language-selective. We discuss evidence that this language network plausibly stores language knowledge and supports core linguistic computations related to accessing words and constructions from memory and combining them to interpret (decode) or generate (encode) linguistic messages. We emphasize that the language network works closely with, but is distinct from, both lower-level - perceptual and motor - mechanisms and higher-level systems of knowledge and reasoning. The perceptual and motor mechanisms process linguistic signals, but, in contrast to the language network, are sensitive only to these signals' surface properties, not their meanings; the systems of knowledge and reasoning (such as the system that supports social reasoning) are sometimes engaged during language use but are not language-selective. This Review lays a foundation both for in-depth investigations of these different components of the language processing pipeline and for probing inter-component interactions.

Hierarchical dynamic coding coordinates speech comprehension in the brain

Speech comprehension requires the human brain to transform an acoustic waveform into meaning. To do so, the brain generates a hierarchy of features that converts the sensory input into increasingly abstract language properties. However, little is known about how these hierarchical features are generated and continuously coordinated. Here, we propose that each linguistic feature is dynamically represented in the brain to simultaneously represent successive events. To test this 'Hierarchical Dynamic Coding' (HDC) hypothesis, we use time-resolved decoding of brain activity to track the construction, maintenance, and integration of a comprehensive hierarchy of language features spanning acoustic, phonetic, sub-lexical, lexical, syntactic and semantic representations. For this, we recorded 21 participants with magnetoencephalography (MEG), while they listened to two hours of short stories. Our analyses reveal three main findings. First, the brain incrementally represents and simultaneously maintains successive features. Second, the duration of these representations depend on their level in the language hierarchy. Third, each representation is maintained by a dynamic neural code, which evolves at a speed commensurate with its corresponding linguistic level. This HDC preserves the maintenance of information over time while limiting the interference between successive features. Overall, HDC reveals how the human brain continuously builds and maintains a language hierarchy during natural speech comprehension, thereby anchoring linguistic theories to their biological implementations.

Natural language processing models reveal neural dynamics of human conversation

Through conversation, humans relay complex information through the alternation of speech production and comprehension. The neural mechanisms that underlie these complementary processes or through which information is precisely conveyed by language, however, remain poorly understood. Here, we used pretrained deep learning natural language processing models in combination with intracranial neuronal recordings to discover neural signals that reliably reflect speech production, comprehension, and their transitions during natural conversation between individuals. Our findings indicate that neural activities that encoded linguistic information were broadly distributed throughout frontotemporal areas across multiple frequency bands. We also find that these activities were specific to the words and sentences being conveyed and that they were dependent on the word's specific context and order. Finally, we demonstrate that these neural patterns partially overlapped during language production and comprehension and that listener-speaker transitions were associated with specific, time-aligned changes in neural activity. Collectively, our findings reveal a dynamical organization of neural activities that subserve language production and comprehension during natural conversation and harness the use of deep learning models in understanding the neural mechanisms underlying human language.

Finding structure during incremental speech comprehension

A core aspect of human speech comprehension is the ability to incrementally integrate consecutive words into a structured and coherent interpretation, aligning with the speaker's intended meaning. This rapid process is subject to multidimensional probabilistic constraints, including both linguistic knowledge and non-linguistic information within specific contexts, and it is their interpretative coherence that drives successful comprehension. To study the neural substrates of this process, we extract word-by-word measures of sentential structure from BERT, a deep language model, which effectively approximates the coherent outcomes of the dynamic interplay among various types of constraints. Using representational similarity analysis, we tested BERT parse depths and relevant corpus-based measures against the spatiotemporally resolved brain activity recorded by electro-/magnetoencephalography when participants were listening to the same sentences. Our results provide a detailed picture of the neurobiological processes involved in the incremental construction of structured interpretations. These findings show when and where coherent interpretations emerge through the evaluation and integration of multifaceted constraints in the brain, which engages bilateral brain regions extending beyond the classical fronto-temporal language system. Furthermore, this study provides empirical evidence supporting the use of artificial neural networks as computational models for revealing the neural dynamics underpinning complex cognitive processes in the brain.

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.

Localizing Syntactic Composition with Left-Corner Recurrent Neural Network Grammars

In computational neurolinguistics, it has been demonstrated that hierarchical models such as recurrent neural network grammars (RNNGs), which jointly generate word sequences and their syntactic structures via the syntactic composition, better explained human brain activity than sequential models such as long short-term memory networks (LSTMs). However, the vanilla RNNG has employed the top-down parsing strategy, which has been pointed out in the psycholinguistics literature as suboptimal especially for head-final/left-branching languages, and alternatively the left-corner parsing strategy has been proposed as the psychologically plausible parsing strategy. In this article, building on this line of inquiry, we investigate not only whether hierarchical models like RNNGs better explain human brain activity than sequential models like LSTMs, but also which parsing strategy is more neurobiologically plausible, by developing a novel fMRI corpus where participants read newspaper articles in a head-final/left-branching language, namely Japanese, through the naturalistic fMRI experiment. The results revealed that left-corner RNNGs outperformed both LSTMs and top-down RNNGs in the left inferior frontal and temporal-parietal regions, suggesting that there are certain brain regions that localize the syntactic composition with the left-corner parsing strategy.

Neurobiological Causal Models of Language Processing

The language faculty is physically realized in the neurobiological infrastructure of the human brain. Despite significant efforts, an integrated understanding of this system remains a formidable challenge. What is missing from most theoretical accounts is a specification of the neural mechanisms that implement language function. Computational models that have been put forward generally lack an explicit neurobiological foundation. We propose a neurobiologically informed causal modeling approach which offers a framework for how to bridge this gap. A neurobiological causal model is a mechanistic description of language processing that is grounded in, and constrained by, the characteristics of the neurobiological substrate. It intends to model the generators of language behavior at the level of implementational causality. We describe key features and neurobiological component parts from which causal models can be built and provide guidelines on how to implement them in model simulations. Then we outline how this approach can shed new light on the core computational machinery for language, the long-term storage of words in the mental lexicon and combinatorial processing in sentence comprehension. In contrast to cognitive theories of behavior, causal models are formulated in the "machine language" of neurobiology which is universal to human cognition. We argue that neurobiological causal modeling should be pursued in addition to existing approaches. Eventually, this approach will allow us to develop an explicit computational neurobiology of language.

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

Many recent studies have shown that representations drawn from neural network language models are extremely effective at predicting brain responses to natural language. But why do these models work so well? One proposed explanation is that language models and brains are similar because they have the same objective: to predict upcoming words before they are perceived. This explanation is attractive because it lends support to the popular theory of predictive coding. We provide several analyses that cast doubt on this claim. First, we show that the ability to predict future words does not uniquely (or even best) explain why some representations are a better match to the brain than others. Second, we show that within a language model, representations that are best at predicting future words are strictly worse brain models than other representations. Finally, we argue in favor of an alternative explanation for the success of language models in neuroscience: These models are effective at predicting brain responses because they generally capture a wide variety of linguistic phenomena.

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.

Dissociable Neural Mechanisms for Human Inference Processing Predicted by Static and Contextual Language Models

Language models (LMs) continue to reveal non-trivial relations to human language performance and the underlying neurophysiology. Recent research has characterized how word embeddings from an LM can be used to generate integrated discourse representations in order to perform inference on events. The current research investigates how such event knowledge may be coded in distinct manners in different classes of LMs and how this maps onto different forms of human inference processing. To do so, we investigate inference on events using two well-documented human experimental protocols from Metusalem et al. (2012) and McKoon and Ratcliff (1986), compared with two protocols for simpler semantic processing. Interestingly, this reveals a dissociation in the relation between local semantics versus event-inference depending on the LM. In a series of experiments, we observed that for the static LMs (word2vec/GloVe), there was a clear dissociation in the relation between semantics and inference for the two inference tasks. In contrast, for the contextual LMs (BERT/RoBERTa), we observed a correlation between semantic and inference processing for both inference tasks. The experimental results suggest that inference as measured by Metusalem and McKoon rely on dissociable processes. While the static models are able to perform Metusalem inference, only the contextual models succeed in McKoon inference. Interestingly, these dissociable processes may be linked to well-characterized automatic versus strategic inference processes in the psychological literature. This allows us to make predictions about dissociable neurophysiological markers that should be found during human inference processing with these tasks.

Alignment of brain embeddings and artificial contextual embeddings in natural language points to common geometric patterns

Contextual embeddings, derived from deep language models (DLMs), provide a continuous vectorial representation of language. This embedding space differs fundamentally from the symbolic representations posited by traditional psycholinguistics. We hypothesize that language areas in the human brain, similar to DLMs, rely on a continuous embedding space to represent language. To test this hypothesis, we densely record the neural activity patterns in the inferior frontal gyrus (IFG) of three participants using dense intracranial arrays while they listened to a 30-minute podcast. From these fine-grained spatiotemporal neural recordings, we derive a continuous vectorial representation for each word (i.e., a brain embedding) in each patient. Using stringent zero-shot mapping we demonstrate that brain embeddings in the IFG and the DLM contextual embedding space have common geometric patterns. The common geometric patterns allow us to predict the brain embedding in IFG of a given left-out word based solely on its geometrical relationship to other non-overlapping words in the podcast. Furthermore, we show that contextual embeddings capture the geometry of IFG embeddings better than static word embeddings. The continuous brain embedding space exposes a vector-based neural code for natural language processing in the human brain.

The cortical representation of language timescales is shared between reading and listening

Language comprehension involves integrating low-level sensory inputs into a hierarchy of increasingly high-level features. Prior work studied brain representations of different levels of the language hierarchy, but has not determined whether these brain representations are shared between written and spoken language. To address this issue, we analyze fMRI BOLD data that were recorded while participants read and listened to the same narratives in each modality. Levels of the language hierarchy are operationalized as timescales, where each timescale refers to a set of spectral components of a language stimulus. Voxelwise encoding models are used to determine where different timescales are represented across the cerebral cortex, for each modality separately. These models reveal that between the two modalities timescale representations are organized similarly across the cortical surface. Our results suggest that, after low-level sensory processing, language integration proceeds similarly regardless of stimulus modality.

Data science opportunities of large language models for neuroscience and biomedicine

Large language models (LLMs) are a new asset class in the machine-learning landscape. Here we offer a primer on defining properties of these modeling techniques. We then reflect on new modes of investigation in which LLMs can be used to reframe classic neuroscience questions to deliver fresh answers. We reason that LLMs have the potential to (1) enrich neuroscience datasets by adding valuable meta-information, such as advanced text sentiment, (2) summarize vast information sources to overcome divides between siloed neuroscience communities, (3) enable previously unthinkable fusion of disparate information sources relevant to the brain, (4) help deconvolve which cognitive concepts most usefully grasp phenomena in the brain, and much more.

Word Frequency and Predictability Dissociate in Naturalistic Reading

Many studies of human language processing have shown that readers slow down at less frequent or less predictable words, but there is debate about whether frequency and predictability effects reflect separable cognitive phenomena: are cognitive operations that retrieve words from the mental lexicon based on sensory cues distinct from those that predict upcoming words based on context? Previous evidence for a frequency-predictability dissociation is mostly based on small samples (both for estimating predictability and frequency and for testing their effects on human behavior), artificial materials (e.g., isolated constructed sentences), and implausible modeling assumptions (discrete-time dynamics, linearity, additivity, constant variance, and invariance over time), which raises the question: do frequency and predictability dissociate in ordinary language comprehension, such as story reading? This study leverages recent progress in open data and computational modeling to address this question at scale. A large collection of naturalistic reading data (six datasets, >2.2 M datapoints) is analyzed using nonlinear continuous-time regression, and frequency and predictability are estimated using statistical language models trained on more data than is currently typical in psycholinguistics. Despite the use of naturalistic data, strong predictability estimates, and flexible regression models, results converge with earlier experimental studies in supporting dissociable and additive frequency and predictability effects.

Large-scale evidence for logarithmic effects of word predictability on reading time

During real-time language comprehension, our minds rapidly decode complex meanings from sequences of words. The difficulty of doing so is known to be related to words' contextual predictability, but what cognitive processes do these predictability effects reflect? In one view, predictability effects reflect facilitation due to anticipatory processing of words that are predictable from context. This view predicts a linear effect of predictability on processing demand. In another view, predictability effects reflect the costs of probabilistic inference over sentence interpretations. This view predicts either a logarithmic or a superlogarithmic effect of predictability on processing demand, depending on whether it assumes pressures toward a uniform distribution of information over time. The empirical record is currently mixed. Here, we revisit this question at scale: We analyze six reading datasets, estimate next-word probabilities with diverse statistical language models, and model reading times using recent advances in nonlinear regression. Results support a logarithmic effect of word predictability on processing difficulty, which favors probabilistic inference as a key component of human language processing.

Driving and suppressing the human language network using large language models

Transformer models such as GPT generate human-like language and are predictive of human brain responses to language. Here, using functional-MRI-measured brain responses to 1,000 diverse sentences, we first show that a GPT-based encoding model can predict the magnitude of the brain response associated with each sentence. We then 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 the 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 non-invasively control neural activity in higher-level cortical areas, such as the language network.

Sentence-level embeddings reveal dissociable word- and sentence-level cortical representation across coarse- and fine-grained levels of meaning

In this large-sample (N = 64) fMRI study, sentence embeddings (text-embedding-ada-002, OpenAI) and representational similarity analysis were used to contrast sentence-level and word-level semantic representation. Overall, sentence-level information resulted in a 20-25 % increase in the model's ability to captures neural representation when compared to word-level only information (word-order scrambled embeddings). This increase was relatively undifferentiated across the cortex. However, when coarse-grained (across thematic category) and fine-grained (within thematic category) combinatorial meaning were separately assessed, word- and sentence-level representations were seen to strongly dissociate across the cortex and to do so differently as a function of grain. Coarse-grained sentence-level representations were evident in occipitotemporal, ventral temporal and medial prefrontal cortex, while fine-grained differences were seen in lateral prefrontal and parietal cortex, middle temporal gyrus, the precuneus, and medial prefrontal cortex. This result indicates dissociable cortical substrates underly single concept versus combinatorial meaning and that different cortical regions specialise for fine- and coarse-grained meaning.

Probing the Representational Structure of Regular Polysemy via Sense Analogy Questions: Insights from Contextual Word Vectors

Regular polysemes are sets of ambiguous words that all share the same relationship between their meanings, such as CHICKEN and LOBSTER both referring to an animal or its meat. To probe how a distributional semantic model, here exemplified by bidirectional encoder representations from transformers (BERT), represents regular polysemy, we analyzed whether its embeddings support answering sense analogy questions similar to "is the mapping between CHICKEN (as an animal) and CHICKEN (as a meat) similar to that which maps between LOBSTER (as an animal) to LOBSTER (as a meat)?" We did so using the LRcos model, which combines a logistic regression classifier of different categories (e.g., animal vs. meat) with a measure of cosine similarity. We found that (a) the model was sensitive to the shared structure within a given regular relationship; (b) the shared structure varies across different regular relationships (e.g., animal/meat vs. location/organization), potentially reflective of a "regularity continuum;" (c) some high-order latent structure is shared across different regular relationships, suggestive of a similar latent structure across different types of relationships; and (d) there is a lack of evidence for the aforementioned effects being explained by meaning overlap. Lastly, we found that both components of the LRcos model made important contributions to accurate responding and that a variation of this method could yield an accuracy boost of 10% in answering sense analogy questions. These findings enrich previous theoretical work on regular polysemy with a computationally explicit theory and methods, and provide evidence for an important organizational principle for the mental lexicon and the broader conceptual knowledge system.

Neural Dynamics of the Processing of Speech Features: Evidence for a Progression of Features from Acoustic to Sentential Processing

When we listen to speech, our brain's neurophysiological responses "track" its acoustic features, but it is less well understood how these auditory responses are modulated by linguistic content. Here, we recorded magnetoencephalography (MEG) responses while subjects listened to four types of continuous-speech-like passages: speech-envelope modulated noise, English-like non-words, scrambled words, and narrative passage. Temporal response function (TRF) analysis provides strong neural evidence for the emergent features of speech processing in cortex, from acoustics to higher-level linguistics, as incremental steps in neural speech processing. Critically, we show a stepwise hierarchical progression of progressively higher order features over time, reflected in both bottom-up (early) and top-down (late) processing stages. Linguistically driven top-down mechanisms take the form of late N400-like responses, suggesting a central role of predictive coding mechanisms at multiple levels. As expected, the neural processing of lower-level acoustic feature responses is bilateral or right lateralized, with left lateralization emerging only for lexical-semantic features. Finally, our results identify potential neural markers of the computations underlying speech perception and comprehension.

Emergence of syntax and word prediction in an artificial neural circuit of the cerebellum

The cerebellum, interconnected with the cerebral neocortex, plays a vital role in human-characteristic cognition such as language processing, however, knowledge about the underlying circuit computation of the cerebellum remains very limited. To gain a better understanding of the computation underlying cerebellar language processing, we developed a biologically constrained cerebellar artificial neural network (cANN) model, which implements the recently identified cerebello-cerebellar recurrent pathway. We found that while cANN acquires prediction of future words, another function of syntactic recognition emerges in the middle layer of the prediction circuit. The recurrent pathway of the cANN was essential for the two language functions, whereas cANN variants with further biological constraints preserved these functions. Considering the uniform structure of cerebellar circuitry across all functional domains, the single-circuit computation, which is the common basis of the two language functions, can be generalized to fundamental cerebellar functions of prediction and grammar-like rule extraction from sequences, that underpin a wide range of cerebellar motor and cognitive functions. This is a pioneering study to understand the circuit computation of human-characteristic cognition using biologically-constrained ANNs.

Short-term Hebbian learning can implement transformer-like attention

Transformers have revolutionized machine learning models of language and vision, but their connection with neuroscience remains tenuous. Built from attention layers, they require a mass comparison of queries and keys that is difficult to perform using traditional neural circuits. Here, we show that neurons can implement attention-like computations using short-term, Hebbian synaptic potentiation. We call our mechanism the match-and-control principle and it proposes that when activity in an axon is synchronous, or matched, with the somatic activity of a neuron that it synapses onto, the synapse can be briefly strongly potentiated, allowing the axon to take over, or control, the activity of the downstream neuron for a short time. In our scheme, the keys and queries are represented as spike trains and comparisons between the two are performed in individual spines allowing for hundreds of key comparisons per query and roughly as many keys and queries as there are neurons in the network.

Information-Restricted Neural Language Models Reveal Different Brain Regions' Sensitivity to Semantics, Syntax, and Context

A fundamental question in neurolinguistics concerns the brain regions involved in syntactic and semantic processing during speech comprehension, both at the lexical (word processing) and supra-lexical levels (sentence and discourse processing). To what extent are these regions separated or intertwined? To address this question, we introduce a novel approach exploiting neural language models to generate high-dimensional feature sets that separately encode semantic and syntactic information. More precisely, we train a lexical language model, GloVe, and a supra-lexical language model, GPT-2, on a text corpus from which we selectively removed either syntactic or semantic information. We then assess to what extent the features derived from these information-restricted models are still able to predict the fMRI time courses of humans listening to naturalistic text. Furthermore, to determine the windows of integration of brain regions involved in supra-lexical processing, we manipulate the size of contextual information provided to GPT-2. The analyses show that, while most brain regions involved in language comprehension are sensitive to both syntactic and semantic features, the relative magnitudes of these effects vary across these regions. Moreover, regions that are best fitted by semantic or syntactic features are more spatially dissociated in the left hemisphere than in the right one, and the right hemisphere shows sensitivity to longer contexts than the left. The novelty of our approach lies in the ability to control for the information encoded in the models' embeddings by manipulating the training set. These "information-restricted" models complement previous studies that used language models to probe the neural bases of language, and shed new light on its spatial organization.

Representational constraints underlying similarity between task-optimized neural systems

Neural systems, artificial and biological, show similar representations of inputs when optimized to perform similar tasks. In visual systems optimized for tasks similar to object recognition, we propose that representation similarities arise from the constraints imposed by the development of abstractions in the representation across the processing stages. To study the effect of abstraction hierarchy of representations across different visual systems, we constructed a two-dimensional space in which each neural representation is positioned based on its distance from the pixel space and the class space. Trajectories of representations in all the task-optimized visual neural networks start close to the pixel space and gradually move towards higher abstract representations, such as object categories. We also observe that proximity in this abstraction space predicts the similarity of neural representations between visual systems. The gradual similar change of the representations suggests that the similarity across different task-optimized systems could arise from constraints on representational trajectories.

The Cortical Representation of Language Timescales is Shared between Reading and Listening

Language comprehension involves integrating low-level sensory inputs into a hierarchy of increasingly high-level features. Prior work studied brain representations of different levels of the language hierarchy, but has not determined whether these brain representations are shared between written and spoken language. To address this issue, we analyzed fMRI BOLD data recorded while participants read and listened to the same narratives in each modality. Levels of the language hierarchy were operationalized as timescales, where each timescale refers to a set of spectral components of a language stimulus. Voxelwise encoding models were used to determine where different timescales are represented across the cerebral cortex, for each modality separately. These models reveal that between the two modalities timescale representations are organized similarly across the cortical surface. Our results suggest that, after low-level sensory processing, language integration proceeds similarly regardless of stimulus modality.

Clarifying status of DNNs as models of human vision

On several key issues we agree with the commentators. Perhaps most importantly, everyone seems to agree that psychology has an important role to play in building better models of human vision, and (most) everyone agrees (including us) that deep neural networks (DNNs) will play an important role in modelling human vision going forward. But there are also disagreements about what models are for, how DNN-human correspondences should be evaluated, the value of alternative modelling approaches, and impact of marketing hype in the literature. In our view, these latter issues are contributing to many unjustified claims regarding DNN-human correspondences in vision and other domains of cognition. We explore all these issues in this response.

Introducing MEG-MASC a high-quality magneto-encephalography dataset for evaluating natural speech processing

The "MEG-MASC" dataset provides a curated set of raw magnetoencephalography (MEG) recordings of 27 English speakers who listened to two hours of naturalistic stories. Each participant performed two identical sessions, involving listening to four fictional stories from the Manually Annotated Sub-Corpus (MASC) intermixed with random word lists and comprehension questions. We time-stamp the onset and offset of each word and phoneme in the metadata of the recording, and organize the dataset according to the 'Brain Imaging Data Structure' (BIDS). This data collection provides a suitable benchmark to large-scale encoding and decoding analyses of temporally-resolved brain responses to speech. We provide the Python code to replicate several validations analyses of the MEG evoked responses such as the temporal decoding of phonetic features and word frequency. All code and MEG, audio and text data are publicly available to keep with best practices in transparent and reproducible research.

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

Models that predict brain responses to stimuli provide one measure of understanding of a sensory system and have many potential applications in science and engineering. Deep artificial neural networks have emerged as the leading such predictive models of the visual system but are less explored in audition. Prior work provided examples of audio-trained neural networks that produced good predictions of auditory cortical fMRI responses and exhibited correspondence between model stages and brain regions, but left it unclear whether these results generalize to other neural network models and, thus, how to further improve models in this domain. We evaluated model-brain correspondence for publicly available audio neural network models along with in-house models trained on 4 different tasks. Most tested models outpredicted standard spectromporal filter-bank models of auditory cortex and exhibited systematic model-brain correspondence: Middle stages best predicted primary auditory cortex, while deep stages best predicted non-primary cortex. However, some state-of-the-art models produced substantially worse brain predictions. Models trained to recognize speech in background noise produced better brain predictions than models trained to recognize speech in quiet, potentially because hearing in noise imposes constraints on biological auditory representations. The training task influenced the prediction quality for specific cortical tuning properties, with best overall predictions resulting from models trained on multiple tasks. The results generally support the promise of deep neural networks as models of audition, though they also indicate that current models do not explain auditory cortical responses in their entirety.

How deep is the brain? The shallow brain hypothesis

Deep learning and predictive coding architectures commonly assume that inference in neural networks is hierarchical. However, largely neglected in deep learning and predictive coding architectures is the neurobiological evidence that all hierarchical cortical areas, higher or lower, project to and receive signals directly from subcortical areas. Given these neuroanatomical facts, today's dominance of cortico-centric, hierarchical architectures in deep learning and predictive coding networks is highly questionable; such architectures are likely to be missing essential computational principles the brain uses. In this Perspective, we present the shallow brain hypothesis: hierarchical cortical processing is integrated with a massively parallel process to which subcortical areas substantially contribute. This shallow architecture exploits the computational capacity of cortical microcircuits and thalamo-cortical loops that are not included in typical hierarchical deep learning and predictive coding networks. We argue that the shallow brain architecture provides several critical benefits over deep hierarchical structures and a more complete depiction of how mammalian brains achieve fast and flexible computational capabilities.

How adults understand what young children say

Children's early speech often bears little resemblance to that of adults, and yet parents and other caregivers are able to interpret that speech and react accordingly. Here we investigate how adult listeners' inferences reflect sophisticated beliefs about what children are trying to communicate, as well as how children are likely to pronounce words. Using a Bayesian framework for modelling spoken word recognition, we find that computational models can replicate adult interpretations of children's speech only when they include strong, context-specific prior expectations about the messages that children will want to communicate. This points to a critical role of adult cognitive processes in supporting early communication and reveals how children can actively prompt adults to take actions on their behalf even when they have only a nascent understanding of the adult language. We discuss the wide-ranging implications of the powerful listening capabilities of adults for theories of first language acquisition.

Dissecting neural computations in the human auditory pathway using deep neural networks for speech

The human auditory system extracts rich linguistic abstractions from speech signals. Traditional approaches to understanding this complex process have used linear feature-encoding models, with limited success. Artificial neural networks excel in speech recognition tasks and offer promising computational models of speech processing. We used speech representations in state-of-the-art deep neural network (DNN) models to investigate neural coding from the auditory nerve to the speech cortex. Representations in hierarchical layers of the DNN correlated well with the neural activity throughout the ascending auditory system. Unsupervised speech models performed at least as well as other purely supervised or fine-tuned models. Deeper DNN layers were better correlated with the neural activity in the higher-order auditory cortex, with computations aligned with phonemic and syllabic structures in speech. Accordingly, DNN models trained on either English or Mandarin predicted cortical responses in native speakers of each language. These results reveal convergence between DNN model representations and the biological auditory pathway, offering new approaches for modeling neural coding in the auditory cortex.

Large language models transition from integrating across position-yoked, exponential windows to structure-yoked, power-law windows

Modern language models excel at integrating across long temporal scales needed to encode linguistic meaning and show non-trivial similarities to biological neural systems. Prior work suggests that human brain responses to language exhibit hierarchically organized "integration windows" that substantially constrain the overall influence of an input token (e.g., a word) on the neural response. However, little prior work has attempted to use integration windows to characterize computations in large language models (LLMs). We developed a simple word-swap procedure for estimating integration windows from black-box language models that does not depend on access to gradients or knowledge of the model architecture (e.g., attention weights). Using this method, we show that trained LLMs exhibit stereotyped integration windows that are well-fit by a convex combination of an exponential and a power-law function, with a partial transition from exponential to power-law dynamics across network layers. We then introduce a metric for quantifying the extent to which these integration windows vary with structural boundaries (e.g., sentence boundaries), and using this metric, we show that integration windows become increasingly yoked to structure at later network layers. None of these findings were observed in an untrained model, which as expected integrated uniformly across its input. These results suggest that LLMs learn to integrate information in natural language using a stereotyped pattern: integrating across position-yoked, exponential windows at early layers, followed by structure-yoked, power-law windows at later layers. The methods we describe in this paper provide a general-purpose toolkit for understanding temporal integration in language models, facilitating cross-disciplinary research at the intersection of biological and artificial intelligence.

Modeling Brain Representations of Words' Concreteness in Context Using GPT-2 and Human Ratings

The meaning of most words in language depends on their context. Understanding how the human brain extracts contextualized meaning, and identifying where in the brain this takes place, remain important scientific challenges. But technological and computational advances in neuroscience and artificial intelligence now provide unprecedented opportunities to study the human brain in action as language is read and understood. Recent contextualized language models seem to be able to capture homonymic meaning variation ("bat", in a baseball vs. a vampire context), as well as more nuanced differences of meaning-for example, polysemous words such as "book", which can be interpreted in distinct but related senses ("explain a book", information, vs. "open a book", object) whose differences are fine-grained. We study these subtle differences in lexical meaning along the concrete/abstract dimension, as they are triggered by verb-noun semantic composition. We analyze functional magnetic resonance imaging (fMRI) activations elicited by Italian verb phrases containing nouns whose interpretation is affected by the verb to different degrees. By using a contextualized language model and human concreteness ratings, we shed light on where in the brain such fine-grained meaning variation takes place and how it is coded. Our results show that phrase concreteness judgments and the contextualized model can predict BOLD activation associated with semantic composition within the language network. Importantly, representations derived from a complex, nonlinear composition process consistently outperform simpler composition approaches. This is compatible with a holistic view of semantic composition in the brain, where semantic representations are modified by the process of composition itself. When looking at individual brain areas, we find that encoding performance is statistically significant, although with differing patterns of results, suggesting differential involvement, in the posterior superior temporal sulcus, inferior frontal gyrus and anterior temporal lobe, and in motor areas previously associated with processing of concreteness/abstractness.

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.

Human attention during goal-directed reading comprehension relies on task optimization

The computational principles underlying attention allocation in complex goal-directed tasks remain elusive. Goal-directed reading, that is, reading a passage to answer a question in mind, is a common real-world task that strongly engages attention. Here, we investigate what computational models can explain attention distribution in this complex task. We show that the reading time on each word is predicted by the attention weights in transformer-based deep neural networks (DNNs) optimized to perform the same reading task. Eye tracking further reveals that readers separately attend to basic text features and question-relevant information during first-pass reading and rereading, respectively. Similarly, text features and question relevance separately modulate attention weights in shallow and deep DNN layers. Furthermore, when readers scan a passage without a question in mind, their reading time is predicted by DNNs optimized for a word prediction task. Therefore, we offer a computational account of how task optimization modulates attention distribution during real-world reading.

EEG-based speaker-listener neural coupling reflects speech-selective attentional mechanisms beyond the speech stimulus

When we pay attention to someone, do we focus only on the sound they make, the word they use, or do we form a mental space shared with the speaker we want to pay attention to? Some would argue that the human language is no other than a simple signal, but others claim that human beings understand each other because they form a shared mental ground between the speaker and the listener. Our study aimed to explore the neural mechanisms of speech-selective attention by investigating the electroencephalogram-based neural coupling between the speaker and the listener in a cocktail party paradigm. The temporal response function method was employed to reveal how the listener was coupled to the speaker at the neural level. The results showed that the neural coupling between the listener and the attended speaker peaked 5 s before speech onset at the delta band over the left frontal region, and was correlated with speech comprehension performance. In contrast, the attentional processing of speech acoustics and semantics occurred primarily at a later stage after speech onset and was not significantly correlated with comprehension performance. These findings suggest a predictive mechanism to achieve speaker-listener neural coupling for successful speech comprehension.

Event Knowledge in Large Language Models: The Gap Between the Impossible and the Unlikely

Word co-occurrence patterns in language corpora contain a surprising amount of conceptual knowledge. Large language models (LLMs), trained to predict words in context, leverage these patterns to achieve impressive performance on diverse semantic tasks requiring world knowledge. An important but understudied question about LLMs' semantic abilities is whether they acquire generalized knowledge of common events. Here, we test whether five pretrained LLMs (from 2018's BERT to 2023's MPT) assign a higher likelihood to plausible descriptions of agent-patient interactions than to minimally different implausible versions of the same event. Using three curated sets of minimal sentence pairs (total n = 1215), we found that pretrained LLMs possess substantial event knowledge, outperforming other distributional language models. In particular, they almost always assign a higher likelihood to possible versus impossible events (The teacher bought the laptop vs. The laptop bought the teacher). However, LLMs show less consistent preferences for likely versus unlikely events (The nanny tutored the boy vs. The boy tutored the nanny). In follow-up analyses, we show that (i) LLM scores are driven by both plausibility and surface-level sentence features, (ii) LLM scores generalize well across syntactic variants (active vs. passive constructions) but less well across semantic variants (synonymous sentences), (iii) some LLM errors mirror human judgment ambiguity, and (iv) sentence plausibility serves as an organizing dimension in internal LLM representations. Overall, our results show that important aspects of event knowledge naturally emerge from distributional linguistic patterns, but also highlight a gap between representations of possible/impossible and likely/unlikely events.

What are large language models supposed to model?

Do large language models (LLMs) constitute a computational account of how humans process language? And if so, what is the role of (psycho)linguistic theory in understanding the relationship between artificial and biological minds? The answer depends on choosing among several, fundamentally distinct ways of interpreting these models as hypotheses about humans.

Neurobiological mechanisms for language, symbols and concepts: Clues from brain-constrained deep neural networks

Neural networks are successfully used to imitate and model cognitive processes. However, to provide clues about the neurobiological mechanisms enabling human cognition, these models need to mimic the structure and function of real brains. Brain-constrained networks differ from classic neural networks by implementing brain similarities at different scales, ranging from the micro- and mesoscopic levels of neuronal function, local neuronal links and circuit interaction to large-scale anatomical structure and between-area connectivity. This review shows how brain-constrained neural networks can be applied to study in silico the formation of mechanisms for symbol and concept processing and to work towards neurobiological explanations of specifically human cognitive abilities. These include verbal working memory and learning of large vocabularies of symbols, semantic binding carried by specific areas of cortex, attention focusing and modulation driven by symbol type, and the acquisition of concrete and abstract concepts partly influenced by symbols. Neuronal assembly activity in the networks is analyzed to deliver putative mechanistic correlates of higher cognitive processes and to develop candidate explanations founded in established neurobiological principles.

Prediction during language comprehension: what is next?

Prediction is often regarded as an integral aspect of incremental language comprehension, but little is known about the cognitive architectures and mechanisms that support it. We review studies showing that listeners and readers use all manner of contextual information to generate multifaceted predictions about upcoming input. The nature of these predictions may vary between individuals owing to differences in language experience, among other factors. We then turn to unresolved questions which may guide the search for the underlying mechanisms. (i) Is prediction essential to language processing or an optional strategy? (ii) Are predictions generated from within the language system or by domain-general processes? (iii) What is the relationship between prediction and memory? (iv) Does prediction in comprehension require simulation via the production system? We discuss promising directions for making progress in answering these questions and for developing a mechanistic understanding of prediction in language.