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.

Narrative thought and decision-making.

A significant body of literature has identified how narrative provides a basis for perceiving and understanding human experience. In the target article, the authors arrive at the need for a form of narrative-based reasoning due to constraints that render probabilistic-based reasoning ineffective. This commentary attempts to bridge this gap and identify links between the proposed and existing theories.

Integration of velocity-dependent spatio-temporal structure of place cell activation during navigation in a reservoir model of prefrontal cortex.

Sequential behavior unfolds both in space and in time. The same spatial trajectory can be realized in different manners in the same overall time by changing instantaneous speeds. The current research investigates how speed profiles might be given behavioral significance and how cortical networks might encode this information. We first demonstrate that rats can associate different speed patterns on the same trajectory with distinct behavioral choices. In this novel experimental paradigm, rats follow a small baited robot in a large megaspace environment where the rat's speed is precisely controlled by the robot's speed. Based on this proof of concept and research showing that recurrent reservoir networks are ideal for representing spatio-temporal structures, we then test reservoir networks in simulated navigation contexts and demonstrate they can discriminate between traversals of the same path with identical durations but different speed profiles. We then test the networks in an embodied robotic setup, where we use place cell representations from physically navigating robots as input and again successfully discriminate between traversals. To demonstrate that this capability is inherent to recurrent networks, we compared the model against simple linear integrators. Interestingly, although the linear integrators could also perform the speed profile discrimination, a clear difference emerged when examining information coding in both models. Reservoir neurons displayed a form of statistical mixed selectivity as a complex interaction between spatial location and speed that was not as abundant in the linear integrators. This mixed selectivity is characteristic of cortex and reservoirs and allows us to generate specific predictions about the neural activity that will be recorded in rat cortex in future experiments.

Narrative event segmentation in the cortical reservoir.

Recent research has revealed that during continuous perception of movies or stories, humans display cortical activity patterns that reveal hierarchical segmentation of event structure. Thus, sensory areas like auditory cortex display high frequency segmentation related to the stimulus, while semantic areas like posterior middle cortex display a lower frequency segmentation related to transitions between events. These hierarchical levels of segmentation are associated with different time constants for processing. Likewise, when two groups of participants heard the same sentence in a narrative, preceded by different contexts, neural responses for the groups were initially different and then gradually aligned. The time constant for alignment followed the segmentation hierarchy: sensory cortices aligned most quickly, followed by mid-level regions, while some higher-order cortical regions took more than 10 seconds to align. These hierarchical segmentation phenomena can be considered in the context of processing related to comprehension. In a recently described model of discourse comprehension word meanings are modeled by a language model pre-trained on a billion word corpus. During discourse comprehension, word meanings are continuously integrated in a recurrent cortical network. The model demonstrates novel discourse and inference processing, in part because of two fundamental characteristics: real-world event semantics are represented in the word embeddings, and these are integrated in a reservoir network which has an inherent gradient of functional time constants due to the recurrent connections. Here we demonstrate how this model displays hierarchical narrative event segmentation properties beyond the embeddings alone, or their linear integration. The reservoir produces activation patterns that are segmented by a hidden Markov model (HMM) in a manner that is comparable to that of humans. Context construction displays a continuum of time constants across reservoir neuron subsets, while context forgetting has a fixed time constant across these subsets. Importantly, virtual areas formed by subgroups of reservoir neurons with faster time constants segmented with shorter events, while those with longer time constants preferred longer events. This neurocomputational recurrent neural network simulates narrative event processing as revealed by the fMRI event segmentation algorithm provides a novel explanation of the asymmetry in narrative forgetting and construction. The model extends the characterization of online integration processes in discourse to more extended narrative, and demonstrates how reservoir computing provides a useful model of cortical processing of narrative structure.

Common ERP responses to narrative incoherence in sentence and picture pair comprehension.

Understanding the neural processes underlying the comprehension of visual images and sentences remains a major open challenge in cognitive neuroscience. We previously demonstrated with fMRI and DTI that comprehension of visual images and sentences describing human activities recruits a common extended parietal-temporal-frontal semantic system. The current research tests the hypothesis that this common semantic system will display similar ERP profiles during processing in these two modalities, providing further support for the common comprehension system. We recorded EEG from naïve subjects as they saw simple narratives made up of a first visual image depicting a human event, followed by a second image that was either a sequentially coherent narrative follow-up, or not, of the first. Incoherent second stimuli depict the same agents but shifted into a different situation. In separate blocks of trials the same protocol was presented using narrative sentence stimuli. Part of the novelty is the comparison of sentence and visual narrative responses. ERPs revealed common neural profiles for narrative processing across image and sentence modalities in the form of early and late central and frontal positivities in response to narrative incoherence. There was an additional posterior positivity only for sentences in a very late window. These results are discussed in the context of ERP signatures of narrative processing and meaning, and a current model of narrative comprehension.

Learning to Use Narrative Function Words for the Organization and Communication of Experience.

How do people learn to talk about the causal and temporal relations between events, and the motivation behind why people do what they do? The narrative practice hypothesis of Hutto and Gallagher holds that children are exposed to narratives that provide training for understanding and expressing reasons for why people behave as they do. In this context, we have recently developed a model of narrative processing where a structured model of the developing situation (the situation model) is built up from experienced events, and enriched by sentences in a narrative that describe event meanings. The main interest is to develop a proof of concept for how narrative can be used to structure, organize and describe experience. Narrative sentences describe events, and they also define temporal and causal relations between events. These relations are specified by a class of narrative function words, including "because, before, after, first, finally." The current research develops a proof of concept that by observing how people describe social events, a developmental robotic system can begin to acquire early knowledge of how to explain the reasons for events. We collect data from naïve subjects who use narrative function words to describe simple scenes of human-robot interaction, and then employ algorithms for extracting the statistical structure of how narrative function words link events in the situation model. By using these statistical regularities, the robot can thus learn from human experience about how to properly employ in question-answering dialogues with the human, and in generating canonical narratives for new experiences. The behavior of the system is demonstrated over several behavioral interactions, and associated narrative interaction sessions, while a more formal extended evaluation and user study will be the subject of future research. Clearly this is far removed from the power of the full blown narrative practice capability, but it provides a first step in the development of an experimental infrastructure for the study of socially situated narrative practice in human-robot interaction.

A Model of Online Temporal-Spatial Integration for Immediacy and Overrule in Discourse Comprehension.

During discourse comprehension, information from prior processing is integrated and appears to be immediately accessible. This was remarkably demonstrated by an N400 for "salted" and not "in love" in response to "The peanut was salted/in love." Discourse overrule was induced by prior discourse featuring the peanut as an animate agent. Immediate discourse overrule requires a model that integrates information at two timescales. One is over the lifetime and includes event knowledge and word semantics. The second is over the discourse in an event context. We propose a model where both are accounted for by temporal-to-spatial integration of experience into distributed spatial representations, providing immediate access to experience accumulated over different timescales. For lexical semantics, this is modeled by a word embedding system trained by sequential exposure to the entire Wikipedia corpus. For discourse, this is modeled by a recurrent reservoir network trained to generate a discourse vector for input sequences of words. The N400 is modeled as the difference between the instantaneous discourse vector and the target word. We predict this model can account for semantic immediacy and discourse overrule. The model simulates lexical priming and discourse overrule in the "Peanut in love" discourse, and it demonstrates that an unexpected word elicits reduced N400 if it is generally related to the event described in prior discourse, and that this effect disappears when the discourse context is removed. This neurocomputational model is the first to simulate immediacy and overrule in discourse-modulated N400, and contributes to characterization of online integration processes in discourse.

Real-time sensory-motor integration of hippocampal place cell replay and prefrontal sequence learning in simulated and physical rat robots for novel path optimization.

An open problem in the cognitive dimensions of navigation concerns how previous exploratory experience is reorganized in order to allow the creation of novel efficient navigation trajectories. This behavior is revealed in the "traveling salesrat problem" (TSP) when rats discover the shortest path linking baited food wells after a few exploratory traversals. We have recently published a model of navigation sequence learning, where sharp wave ripple replay of hippocampal place cells transmit "snippets" of the recent trajectories that the animal has explored to the prefrontal cortex (PFC) (Cazin et al. in PLoS Comput Biol 15:e1006624, 2019). PFC is modeled as a recurrent reservoir network that is able to assemble these snippets into the efficient sequence (trajectory of spatial locations coded by place cell activation). The model of hippocampal replay generates a distribution of snippets as a function of their proximity to a reward, thus implementing a form of spatial credit assignment that solves the TSP task. The integrative PFC reservoir reconstructs the efficient TSP sequence based on exposure to this distribution of snippets that favors paths that are most proximal to rewards. While this demonstrates the theoretical feasibility of the PFC-HIPP interaction, the integration of such a dynamic system into a real-time sensory-motor system remains a challenge. In the current research, we test the hypothesis that the PFC reservoir model can operate in a real-time sensory-motor loop. Thus, the main goal of the paper is to validate the model in simulated and real robot scenarios. Place cell activation encoding the current position of the simulated and physical rat robot feeds the PFC reservoir which generates the successor place cell activation that represents the next step in the reproduced sequence in the readout. This is input to the robot, which advances to the coded location and then generates de novo the current place cell activation. This allows demonstration of the crucial role of embodiment. If the spatial code readout from PFC is played back directly into PFC, error can accumulate, and the system can diverge from desired trajectories. This required a spatial filter to decode the PFC code to a location and then recode a new place cell code for that location. In the robot, the place cell vector output of PFC is used to physically displace the robot and then generate a new place cell coded input to the PFC, replacing part of the software recoding procedure that was required otherwise. We demonstrate how this integrated sensory-motor system can learn simple navigation sequences and then, importantly, how it can synthesize novel efficient sequences based on prior experience, as previously demonstrated (Cazin et al. 2019). This contributes to the understanding of hippocampal replay in novel navigation sequence formation and the important role of embodiment.

Common Neural System for Sentence and Picture Comprehension Across Languages: A Chinese-Japanese Bilingual Study.

While common semantic representations for individual words across languages have been identified, a common meaning system at sentence-level has not been determined. In this study, fMRI was used to investigate whether an across-language sentence comprehension system exists. Chinese-Japanese bilingual participants (n = 32) were asked to determine whether two consecutive stimuli were related (coherent) or not (incoherent) to the same event. Stimuli were displayed with three different modalities (Chinese written sentences, Japanese written sentences, and pictures). The behavioral results showed no significant difference in accuracy and response times among the three modalities. Multi-voxel pattern analysis (MVPA) of fMRI data was used to classify the semantic relationship (coherent or incoherent) across the stimulus modalities. The classifier was first trained to determine coherency within Chinese sentences and then tested with Japanese sentences, and vice versa. A whole-brain searchlight analysis revealed significant above-chance classification accuracy across Chinese and Japanese sentences in the supramarginal gyrus (BA 40), extending into the angular gyrus (BA 39) as well as the opercular (BA 44) and triangular (BA 45) parts of the inferior frontal gyrus in the left hemisphere (cluster-level FWE corrected p < 0.05). Significant above-chance classification accuracy was also found across Japanese sentences and pictures in the supramarginal (BA 40) and angular gyrus (BA 39). These results indicate that a common meaning system for sentence processing across languages and modalities exists, and it involves the left inferior parietal gyrus.

Language is not a gadget.

Heyes does well to argue that some of the apparently innate human capabilities for cultural learning can be considered in terms of more general-purpose mechanisms. In the application of this to language, she overlooks some of its most interesting properties. I review three, and then illustrate how mindreading can come from general-purpose mechanism via language.

Reservoir computing model of prefrontal cortex creates novel combinations of previous navigation sequences from hippocampal place-cell replay with spatial reward propagation.

As rats learn to search for multiple sources of food or water in a complex environment, they generate increasingly efficient trajectories between reward sites. Such spatial navigation capacity involves the replay of hippocampal place-cells during awake states, generating small sequences of spatially related place-cell activity that we call "snippets". These snippets occur primarily during sharp-wave-ripples (SWRs). Here we focus on the role of such replay events, as the animal is learning a traveling salesperson task (TSP) across multiple trials. We hypothesize that snippet replay generates synthetic data that can substantially expand and restructure the experience available and make learning more optimal. We developed a model of snippet generation that is modulated by reward, propagated in the forward and reverse directions. This implements a form of spatial credit assignment for reinforcement learning. We use a biologically motivated computational framework known as 'reservoir computing' to model prefrontal cortex (PFC) in sequence learning, in which large pools of prewired neural elements process information dynamically through reverberations. This PFC model consolidates snippets into larger spatial sequences that may be later recalled by subsets of the original sequences. Our simulation experiments provide neurophysiological explanations for two pertinent observations related to navigation. Reward modulation allows the system to reject non-optimal segments of experienced trajectories, and reverse replay allows the system to "learn" trajectories that it has not physically experienced, both of which significantly contribute to the TSP behavior.

Grammatical verb aspect and event roles in sentence processing.

Two experiments examine how grammatical verb aspect constrains our understanding of events. According to linguistic theory, an event described in the perfect aspect (John had opened the bottle) should evoke a mental representation of a finished event with focus on the resulting object, whereas an event described in the imperfective aspect (John was opening the bottle) should evoke a representation of the event as ongoing, including all stages of the event, and focusing all entities relevant to the ongoing action (instruments, objects, agents, locations, etc.). To test this idea, participants saw rebus sentences in the perfect and imperfective aspect, presented one word at a time, self-paced. In each sentence, the instrument and the recipient of the action were replaced by pictures (John was using/had used a *corkscrew* to open the *bottle* at the restaurant). Time to process the two images as well as speed and accuracy on sensibility judgments were measured. Although experimental sentences always made sense, half of the object and instrument pictures did not match the temporal constraints of the verb. For instance, in perfect sentences aspect-congruent trials presented an image of the corkscrew closed (no longer in-use) and the wine bottle fully open. The aspect-incongruent yet still sensible versions either replaced the corkscrew with an in-use corkscrew (open, in-hand) or the bottle image with a half-opened bottle. In this case, the participant would still respond "yes", but with longer expected response times. A three-way interaction among Verb Aspect, Sentence Role, and Temporal Match on image processing times showed that participants were faster to process images that matched rather than mismatched the aspect of the verb, especially for resulting objects in perfect sentences. A second experiment replicated and extended the results to confirm that this was not due to the placement of the object in the sentence. These two experiments extend previous research, showing how verb aspect drives not only the temporal structure of event representation, but also the focus on specific roles of the event. More generally, the findings of visual match during online sentence-picture processing are consistent with theories of perceptual simulation.

The Role of Autobiographical Memory in the Development of a Robot Self.

This article briefly reviews research in cognitive development concerning the nature of the human self. It then reviews research in developmental robotics that has attempted to retrace parts of the developmental trajectory of the self. This should be of interest to developmental psychologists, and researchers in developmental robotics. As a point of departure, one of the most characteristic aspects of human social interaction is cooperation-the process of entering into a joint enterprise to achieve a common goal. Fundamental to this ability to cooperate is the underlying ability to enter into, and engage in, a self-other relation. This suggests that if we intend for robots to cooperate with humans, then to some extent robots must engage in these self-other relations, and hence they must have some aspect of a self. Decades of research in human cognitive development indicate that the self is not fully present from the outset, but rather that it is developed in a usage-based fashion, that is, through engaging with the world, including the physical world and the social world of animate intentional agents. In an effort to characterize the self, Ulric Neisser noted that self is not unitary, and he thus proposed five types of self-knowledge that correspond to five distinct components of self: ecological, interpersonal, conceptual, temporally extended, and private. He emphasized the ecological nature of each of these levels, how they are developed through the engagement of the developing child with the physical and interpersonal worlds. Crucially, development of the self has been shown to rely on the child's autobiographical memory. From the developmental robotics perspective, this suggests that in principal it would be possible to develop certain aspects of self in a robot cognitive system where the robot is engaged in the physical and social world, equipped with an autobiographical memory system. We review a series of developmental robotics studies that make progress in this enterprise. We conclude with a summary of the properties that are required for the development of these different levels of self, and we identify topics for future research.

Reservoir Computing Properties of Neural Dynamics in Prefrontal Cortex.

Primates display a remarkable ability to adapt to novel situations. Determining what is most pertinent in these situations is not always possible based only on the current sensory inputs, and often also depends on recent inputs and behavioral outputs that contribute to internal states. Thus, one can ask how cortical dynamics generate representations of these complex situations. It has been observed that mixed selectivity in cortical neurons contributes to represent diverse situations defined by a combination of the current stimuli, and that mixed selectivity is readily obtained in randomly connected recurrent networks. In this context, these reservoir networks reproduce the highly recurrent nature of local cortical connectivity. Recombining present and past inputs, random recurrent networks from the reservoir computing framework generate mixed selectivity which provides pre-coded representations of an essentially universal set of contexts. These representations can then be selectively amplified through learning to solve the task at hand. We thus explored their representational power and dynamical properties after training a reservoir to perform a complex cognitive task initially developed for monkeys. The reservoir model inherently displayed a dynamic form of mixed selectivity, key to the representation of the behavioral context over time. The pre-coded representation of context was amplified by training a feedback neuron to explicitly represent this context, thereby reproducing the effect of learning and allowing the model to perform more robustly. This second version of the model demonstrates how a hybrid dynamical regime combining spatio-temporal processing of reservoirs, and input driven attracting dynamics generated by the feedback neuron, can be used to solve a complex cognitive task. We compared reservoir activity to neural activity of dorsal anterior cingulate cortex of monkeys which revealed similar network dynamics. We argue that reservoir computing is a pertinent framework to model local cortical dynamics and their contribution to higher cognitive function.

Sequential Coherence in Sentence Pairs Enhances Imagery during Comprehension: An Individual Differences Study.

The present study investigates how sequential coherence in sentence pairs (events in sequence vs. unrelated events) affects the perceived ability to form a mental image of the sentences for both auditory and visual presentations. In addition, we investigated how the ease of event imagery affected online comprehension (word reading times) in the case of sequentially coherent and incoherent sentence pairs. Two groups of comprehenders were identified based on their self-reported ability to form vivid mental images of described events. Imageability ratings were higher and faster for pairs of sentences that described events in coherent sequences rather than non-sequential events, especially for high imagers. Furthermore, reading times on individual words suggested different comprehension patterns with respect to sequence coherence for the two groups of imagers, with high imagers activating richer mental images earlier than low imagers. The present results offer a novel link between research on imagery and discourse coherence, with specific contributions to our understanding of comprehension patterns for high and low imagers.

Corticostriatal response selection in sentence production: Insights from neural network simulation with reservoir computing.

Language production requires selection of the appropriate sentence structure to accommodate the communication goal of the speaker - the transmission of a particular meaning. Here we consider event meanings, in terms of predicates and thematic roles, and we address the problem that a given event can be described from multiple perspectives, which poses a problem of response selection. We present a model of response selection in sentence production that is inspired by the primate corticostriatal system. The model is implemented in the context of reservoir computing where the reservoir - a recurrent neural network with fixed connections - corresponds to cortex, and the readout corresponds to the striatum. We demonstrate robust learning, and generalization properties of the model, and demonstrate its cross linguistic capabilities in English and Japanese. The results contribute to the argument that the corticostriatal system plays a role in response selection in language production, and to the stance that reservoir computing is a valid potential model of corticostriatal processing.

Exploring the acquisition and production of grammatical constructions through human-robot interaction with echo state networks.

One of the principal functions of human language is to allow people to coordinate joint action. This includes the description of events, requests for action, and their organization in time. A crucial component of language acquisition is learning the grammatical structures that allow the expression of such complex meaning related to physical events. The current research investigates the learning of grammatical constructions and their temporal organization in the context of human-robot physical interaction with the embodied sensorimotor humanoid platform, the iCub. We demonstrate three noteworthy phenomena. First, a recurrent network model is used in conjunction with this robotic platform to learn the mappings between grammatical forms and predicate-argument representations of meanings related to events, and the robot's execution of these events in time. Second, this learning mechanism functions in the inverse sense, i.e., in a language production mode, where rather than executing commanded actions, the robot will describe the results of human generated actions. Finally, we collect data from naïve subjects who interact with the robot via spoken language, and demonstrate significant learning and generalization results. This allows us to conclude that such a neural language learning system not only helps to characterize and understand some aspects of human language acquisition, but also that it can be useful in adaptive human-robot interaction.

Comparison of classifiers for decoding sensory and cognitive information from prefrontal neuronal populations.

Decoding neuronal information is important in neuroscience, both as a basic means to understand how neuronal activity is related to cerebral function and as a processing stage in driving neuroprosthetic effectors. Here, we compare the readout performance of six commonly used classifiers at decoding two different variables encoded by the spiking activity of the non-human primate frontal eye fields (FEF): the spatial position of a visual cue, and the instructed orientation of the animal's attention. While the first variable is exogenously driven by the environment, the second variable corresponds to the interpretation of the instruction conveyed by the cue; it is endogenously driven and corresponds to the output of internal cognitive operations performed on the visual attributes of the cue. These two variables were decoded using either a regularized optimal linear estimator in its explicit formulation, an optimal linear artificial neural network estimator, a non-linear artificial neural network estimator, a non-linear naïve Bayesian estimator, a non-linear Reservoir recurrent network classifier or a non-linear Support Vector Machine classifier. Our results suggest that endogenous information such as the orientation of attention can be decoded from the FEF with the same accuracy as exogenous visual information. All classifiers did not behave equally in the face of population size and heterogeneity, the available training and testing trials, the subject's behavior and the temporal structure of the variable of interest. In most situations, the regularized optimal linear estimator and the non-linear Support Vector Machine classifiers outperformed the other tested decoders.

Dissociable dorsal and ventral frontostriatal working memory circuits: evidence from subthalamic stimulation in Parkinson's disease.

In this study, we investigated the neural substrates involved in visual working memory (WM) and the resulting effects of subthalamic nucleus (STN) stimulation in Parkinson's disease (PD). Cerebral activation revealed by positron emission tomography was compared among Parkinson patients with (PD-ON) or without (PD-OFF) STN stimulation, and a group of control subjects (CT) in two visual WM tasks with spatial (SP) and nonspatial (NSP) components. PD-OFF patients displayed significant reaction time (RT) deficits for both memory tasks. Although there were no significant differences in RT between patients with PD-ON and -OFF stimulation, patients with PD-ON stimulation performed comparably to controls. The memory tasks were executed with normal error rates in PD-ON and -OFF stimulation. In contrast to these behavioral results, whether the corresponding prefrontal activation was differentially affected by deep brain stimulation status in patients depended on whether the WM modality was SP versus NSP. Thus, SP WM was associated with (1) abnormal reduction in dorsolateral prefrontal activity in PD-OFF and -ON stimulation and (2) abnormal overactivation in parieto-temporal cortex in PD-OFF and in limbic circuits in PD-ON stimulation. In NSP WM, normal activation of the ventral prefrontal cortex was restored in PD-ON stimulation. In both visual modalities the posterior cerebral regions including fusiform cortex and cerebellum, displayed abnormally reduced activity in PD. These results indicate that PD induces a prefrontal hypoactivation that STN stimulation can partially restore in a modality selective manner by additional recruitment of limbic structures in SP WM or by recovery of the ventral prefrontal activation in NSP WM.