Conference Agenda

A fundamental question in understanding how the human brain conducts predictive processing is how information about prediction errors is encoded across different levels of the cortical hierarchy, and how this is modulated by attention. Do multiple signals across neuroanatomical regions encode the same information (redundancy), or is additional information revealed when considering these signals together (synergy)?

To address this, we present results from 18 human patients who listened to an auditory oddball task, either while paying attention to the auditory tones or while performing a unrelated visual task. During both conditions, stereoelectroencephalography (sEEG) recordings were obtained from cortical and sub-cortical regions.

By computing mutual information (MI), we quantified the relationship between two neural markers of prediction errors (event-related potentials; ERPs, and broadband activity), and the stimulus identity. Our results show that the amount of auditory prediction error information encoded in the thalamic and frontal regions is modulated by attentional demands: when attending to the visual task, less PE-information is encoded in both regions.

We further employed co-information (Co-I), a higher-order information-theoretic measure, to decompose neural activity into synergistic and redundant components. This analysis revealed that prediction error information is encoded highly synergistically both within and between different levels of the cortical hierarchy including the thalamus, and temporal areas. Notably, we found that attentional demands selectively modulate the degree of this synergistic encoding.

Based on previous research, we propose that this synergistic encoding reflects recurrent processing between different neuroanatomical areas, and its modulation the effect attention has on this processing.

Using subjective reports to measure consciously experienced sense of control has been criticized for low reliability, leading to the domination of this field by the use of implicit measures: intentional binding and sensory attenuation. However, using implicit methods is not always feasible. In the current study, we systematically investigated how the choice of response scale affects reliability of participants’ reports of sense of control. We designed a computer task in which in each trial we manipulated the degree of objective control that participants had. Immediately after each trial participants reported their experienced sense of control on a scale. Across five experiments we compared responses on five scales: a linear scale, a percentage scale, 7-point Likert, 4-point Likert and a binary scale. Moreover, in two additional experiments we compared the directionality of the linear scale. We found that the results were remarkably consistent across all except the binary scale: participants demonstrated the tendency to underestimate the level of control, especially for intermediate levels of objective control. This tendency was not affected by the directionality of the response scale. Next, we tested the between-block reliability of responses finding high level of reliability of responses from the first block. We also performed cluster analysis revealing three distinctive clusters. Finally, we propose how our task can be used as a measure of individual differences and we present metrics that can be extracted from participants’ responses. Overall, our results suggest that explicit reports can be characterized by high reliability typically unaffected by the scale type.

Growing research interest has been dedicated to the idea that sleep-like activity in the brain promotes disengagement from the external world leading to inattention. In particular, slow waves in wakefulness resembling those during sleep, but smaller in amplitude and extent have been found to precede attentional lapses in neural recordings via elecroenchephalography (EEG). Our previous study (Andrillon et al., 2021) demonstrated that the locations of slow waves can differentiate between different mental states: frontal slow wave activity is associated with increased likelihood of mind wandering and posterior slow waves with mind blanking. To overcome the subpar spatial resolution of EEG, we conducted a simultaneous EEG/fMRI experiment with 25 participants, who completed the Sustained Attention to Response Task (SART) for approximately 50 minutes in an MRI scanner. During the task, the participants were prompted to report their mental state at random intervals. Thus, we were able to combine multimodal neuroimaging data with task performance measures and experience sampling. Our findings confirmed that slow wave activity predicted attentional disengagement indicated by performance impairments and subjective reports. Moreover, we have also identified localised associations between slow waves activity and BOLD activity changes in areas in posterior brain regions. These findings provide new evidence for the location-specific effect of sleep-like slow wave activity during attentional lapses in wakefulness.

Introduction

The notion of a contentful mind is challenged by brief moments of no reportable content, termed mind-blanking (MB). MB's neuronal correlates resemble brain markers of sleep, including an fMRI "hyperconnected" pattern of cortex-wide coherence and EEG slow wave-like activity, which to date have been studies in separation. The present study aimed to bridge fMRI functional connectivity (FC) with EEG to a) replicate MB-related brain patterns during task and b) explore the MB-related brain patterns’ electrophysiological basis.

Methods

Thirty-eight participants performed a sustained attention-to-response task (SART) with simultaneous EEG-fMRI recordings and experience sampling of mental states (Blank, On-Task, Off-Task) and alertness (Very Alert, Alert, Sleepy, Very Sleepy). Phase-based coherence estimated fMRI connectivity, and K-means clustering identified brain patterns. For each TR, FC's distance to these patterns was measured to link cognition and alertness to brain patterns. Canonical correlations were used to assess the relationship between FC and slow wave-like activity.

Results

Behaviorally, MB report frequency increased during lower alertness. Neuronally, five brain patterns emerged, characterized by regional anticorrelations and hyperconnectivity. While sleepiness was closest to brain patterns of hyperconnectivity, MB was associated with patterns of anticorrelations. Finally, increases in the amplitude and the slope of the slow wave-like activity correlated with FC proximity to the hyperconnectivity brain pattern.

Conclusions

Our results show that MB and sleepiness have dissociable cortical correlates. With the hyperconnectivity pattern tied to slow wave-like activity, our study advances discussions on how functional neuroimaging shapes ongoing cognition and how slow-wave activity promotes distinct patterns of cortical organization.

We have relatively little understanding of how and why the neural networks in modern AI systems do what they do. One avenue to better understand these systems is to investigate and develop their capacity to introspect and explain their own functioning. Here, we show that contemporary large language models (LLMs) are capable of accurately explaining their own internal processes. To demonstrate this, we train GPT-4o and GPT-4o mini to learn the preferences of hypothetical actors using examples of the actors’ decisions (e.g., “Macbeth is shopping for a condo. If offered the two described below, he would choose Option A”). Each actor's decisions are determined by a computational model that the LLMs are told nothing about (e.g., Macbeth cares enormously about the amount of natural light, but less about the walkability of the neighborhood, and actually prefers smaller condos). We show that, when asked to make decisions on the actors’ behalf, the LLMs not only do so, but can provide accurate, quantitative descriptions of how much they are weighing each of the different attributes. Critically, these reports are genuine introspection: The LLMs are not doing inference based on observing their own responses. (They make them with no memory of their other responses and without any chain-of-thought.) And they are not just reporting “common sense”. (Versions of the LLMs that do not observe the actors will use and report completely different weights for the attributes.) In ongoing work, we are exploring methods to improve the introspective abilities of LLMs.