A hierarchy of linguistic predictions during natural language comprehension

Multiple Concurrent Predictions Inform Prediction Error in the Human Auditory Pathway

The key assumption of the predictive coding framework is that internal representations are used to generate predictions on how the sensory input will look like in the immediate future. These predictions are tested against the actual input by the so-called prediction error units, which encode the residuals of the predictions. What happens to prediction errors, however, if predictions drawn by different stages of the sensory hierarchy contradict each other? To answer this question, we conducted two fMRI experiments while female and male human participants listened to sequences of sounds: pure tones in the first experiment and frequency-modulated sweeps in the second experiment. In both experiments, we used repetition to induce predictions based on stimulus statistics (stats-informed predictions) and abstract rules disclosed in the task instructions to induce an orthogonal set of (task-informed) predictions. We tested three alternative scenarios: neural responses in the auditory sensory pathway encode prediction error with respect to (1) the stats-informed predictions, (2) the task-informed predictions, or (3) a combination of both. Results showed that neural populations in all recorded regions (bilateral inferior colliculus, medial geniculate body, and primary and secondary auditory cortices) encode prediction error with respect to a combination of the two orthogonal sets of predictions. The findings suggest that predictive coding exploits the non-linear architecture of the auditory pathway for the transmission of predictions. Such non-linear transmission of predictions might be crucial for the predictive coding of complex auditory signals like speech.Significance Statement Sensory systems exploit our subjective expectations to make sense of an overwhelming influx of sensory signals. It is still unclear how expectations at each stage of the processing pipeline are used to predict the representations at the other stages. The current view is that this transmission is hierarchical and linear. Here we measured fMRI responses in auditory cortex, sensory thalamus, and midbrain while we induced two sets of mutually inconsistent expectations on the sensory input, each putatively encoded at a different stage. We show that responses at all stages are concurrently shaped by both sets of expectations. The results challenge the hypothesis that expectations are transmitted linearly and provide for a normative explanation of the non-linear physiology of the corticofugal sensory system.

Bayesian Surprise Predicts Human Event Segmentation in Story Listening

Event segmentation theory posits that people segment continuous experience into discrete events and that event boundaries occur when there are large transient increases in prediction error. Here, we set out to test this theory in the context of story listening, by using a deep learning language model (GPT-2) to compute the predicted probability distribution of the next word, at each point in the story. For three stories, we used the probability distributions generated by GPT-2 to compute the time series of prediction error. We also asked participants to listen to these stories while marking event boundaries. We used regression models to relate the GPT-2 measures to the human segmentation data. We found that event boundaries are associated with transient increases in Bayesian surprise but not with a simpler measure of prediction error (surprisal) that tracks, for each word in the story, how strongly that word was predicted at the previous time point. These results support the hypothesis that prediction error serves as a control mechanism governing event segmentation and point to important differences between operational definitions of prediction error.

A shared linguistic space for transmitting our thoughts from brain to brain in natural conversations

Effective communication hinges on a mutual understanding of word meaning in different contexts. The embedding space learned by large language models can serve as an explicit model of the shared, context-rich meaning space humans use to communicate their thoughts. We recorded brain activity using electrocorticography during spontaneous, face-to-face conversations in five pairs of epilepsy patients. We demonstrate that the linguistic embedding space can capture the linguistic content of word-by-word neural alignment between speaker and listener. Linguistic content emerged in the speaker's brain before word articulation, and the same linguistic content rapidly reemerged in the listener's brain after word articulation. These findings establish a computational framework to study how human brains transmit their thoughts to one another in real-world contexts.

Context-based facilitation of semantic access follows both logarithmic and linear functions of stimulus probability

Stimuli are easier to process when context makes them predictable, but does context-based facilitation arise from preactivation of a limited set of relatively probable upcoming stimuli (with facilitation then linearly related to probability) or, instead, because the system maintains and updates a probability distribution across all items (with facilitation logarithmically related to probability)? We measured the N400, an index of semantic access, to words of varying probability, including unpredictable words. Word predictability was measured using both cloze probabilities and a state-of-the-art machine learning language model (GPT-2). We reanalyzed five datasets (n = 138) to demonstrate and then replicate that context-based facilitation on the N400 is graded, even among unpredictable words. Furthermore, we established that the relationship between word predictability and context-based facilitation combines linear and logarithmic functions. We argue that this composite function reveals properties of the mapping between words and semantic features and how feature- and word-related information is activated on-line.

Shared computational principles for language processing in humans and deep language models

Departing from traditional linguistic models, advances in deep learning have resulted in a new type of predictive (autoregressive) deep language models (DLMs). Using a self-supervised next-word prediction task, these models generate appropriate linguistic responses in a given context. In the current study, nine participants listened to a 30-min podcast while their brain responses were recorded using electrocorticography (ECoG). We provide empirical evidence that the human brain and autoregressive DLMs share three fundamental computational principles as they process the same natural narrative: (1) both are engaged in continuous next-word prediction before word onset; (2) both match their pre-onset predictions to the incoming word to calculate post-onset surprise; (3) both rely on contextual embeddings to represent words in natural contexts. Together, our findings suggest that autoregressive DLMs provide a new and biologically feasible computational framework for studying the neural basis of language.

Deep language models have revolutionized natural language processing. The paper discovers three computational principles shared between deep language models and the human brain, which can transform our understanding of the neural basis of language.

Brains and algorithms partially converge in natural language processing

Deep learning algorithms trained to predict masked words from large amount of text have recently been shown to generate activations similar to those of the human brain. However, what drives this similarity remains currently unknown. Here, we systematically compare a variety of deep language models to identify the computational principles that lead them to generate brain-like representations of sentences. Specifically, we analyze the brain responses to 400 isolated sentences in a large cohort of 102 subjects, each recorded for two hours with functional magnetic resonance imaging (fMRI) and magnetoencephalography (MEG). We then test where and when each of these algorithms maps onto the brain responses. Finally, we estimate how the architecture, training, and performance of these models independently account for the generation of brain-like representations. Our analyses reveal two main findings. First, the similarity between the algorithms and the brain primarily depends on their ability to predict words from context. Second, this similarity reveals the rise and maintenance of perceptual, lexical, and compositional representations within each cortical region. Overall, this study shows that modern language algorithms partially converge towards brain-like solutions, and thus delineates a promising path to unravel the foundations of natural language processing.

Charlotte Caucheteux and Jean-Rémi King examine the ability of transformer neural networks trained on word prediction tasks to fit representations in the human brain measured with fMRI and MEG. Their results provide further insight into the workings of transformer language models and their relevance to brain responses.

Cortical activity during naturalistic music listening reflects short-range predictions based on long-term experience

Expectations shape our experience of music. However, the internal model upon which listeners form melodic expectations is still debated. Do expectations stem from Gestalt-like principles or statistical learning? If the latter, does long-term experience play an important role, or are short-term regularities sufficient? And finally, what length of context informs contextual expectations? To answer these questions, we presented human listeners with diverse naturalistic compositions from Western classical music, while recording neural activity using MEG. We quantified note-level melodic surprise and uncertainty using various computational models of music, including a state-of-the-art transformer neural network. A time-resolved regression analysis revealed that neural activity over fronto-temporal sensors tracked melodic surprise particularly around 200ms and 300-500ms after note onset. This neural surprise response was dissociated from sensory-acoustic and adaptation effects. Neural surprise was best predicted by computational models that incorporated long-term statistical learning-rather than by simple, Gestalt-like principles. Yet, intriguingly, the surprise reflected primarily short-range musical contexts of less than ten notes. We present a full replication of our novel MEG results in an openly available EEG dataset. Together, these results elucidate the internal model that shapes melodic predictions during naturalistic music listening.

Language Dysfunction in Schizophrenia: Assessing Neural Tracking to Characterize the Underlying Disorder(s)?

Deficits in language production and comprehension are characteristic of schizophrenia. To date, it remains unclear whether these deficits arise from dysfunctional linguistic knowledge, or dysfunctional predictions derived from the linguistic context. Alternatively, the deficits could be a result of dysfunctional neural tracking of auditory information resulting in decreased auditory information fidelity and even distorted information. Here, we discuss possible ways for clinical neuroscientists to employ neural tracking methodology to independently characterize deficiencies on the auditory-sensory and abstract linguistic levels. This might lead to a mechanistic understanding of the deficits underlying language related disorder(s) in schizophrenia. We propose to combine naturalistic stimulation, measures of speech-brain synchronization, and computational modeling of abstract linguistic knowledge and predictions. These independent but likely interacting assessments may be exploited for an objective and differential diagnosis of schizophrenia, as well as a better understanding of the disorder on the functional level-illustrating the potential of neural tracking methodology as translational tool in a range of psychotic populations.