The cognitive reality monitoring network and theories of consciousness.

Theories of consciousness abound. However, it is difficult to arbitrate reliably among competing theories because they target different levels of neural and cognitive processing or anatomical loci, and only some were developed with computational models in mind. In particular, theories of consciousness need to fully address the three levels of understanding of the brain proposed by David Marr: computational theory, algorithms and hardware. Most major theories refer to only one or two levels, often indirectly. The cognitive reality monitoring network (CRMN) model is derived from computational theories of mixture-of-experts architecture, hierarchical reinforcement learning and generative/inference computing modules, addressing all three levels of understanding. A central feature of the CRMN is the mapping of a gating network onto the prefrontal cortex, making it a prime coding circuit involved in monitoring the accuracy of one's mental states and distinguishing them from external reality. Because the CRMN builds on the hierarchical and layer structure of the cerebral cortex, it may connect research and findings across species, further enabling concrete computational models of consciousness with new, explicitly testable hypotheses. In sum, we discuss how the CRMN model can help further our understanding of the nature and function of consciousness.

Goals, usefulness and abstraction in value-based choice.

Colombian drug lord Pablo Escobar, while on the run, purportedly burned two million dollars in banknotes to keep his daughter warm. A stark reminder that, in life, circumstances and goals can quickly change, forcing us to reassess and modify our values on-the-fly. Studies in decision-making and neuroeconomics have often implicitly equated value to reward, emphasising the hedonic and automatic aspect of the value computation, while overlooking its functional (concept-like) nature. Here we outline the computational and biological principles that enable the brain to compute the usefulness of an option or action by creating abstractions that flexibly adapt to changing goals. We present different algorithmic architectures, comparing ideas from artificial intelligence (AI) and cognitive neuroscience with psychological theories and, when possible, drawing parallels.

Generalized attention-weighted reinforcement learning.

In neuroscience, attention has been shown to bidirectionally interact with reinforcement learning (RL) to reduce the dimensionality of task representations, restricting computations to relevant features. In machine learning, despite their popularity, attention mechanisms have seldom been administered to decision-making problems. Here, we leverage a theoretical model from computational neuroscience - the attention-weighted RL (AWRL), defining how humans identify task-relevant features (i.e., that allow value predictions) - to design an applied deep RL paradigm. We formally demonstrate that the conjunction of the self-attention mechanism, widely employed in machine learning, with value function approximation is a general formulation of the AWRL model. To evaluate our agent, we train it on three Atari tasks at different complexity levels, incorporating both task-relevant and irrelevant features. Because the model uses semantic observations, we can uncover not only which features the agent elects to base decisions on, but also how it chooses to compile more complex, relational features from simpler ones. We first show that performance depends in large part on the ability to compile new compound features, rather than mere focus on individual features. In line with neuroscience predictions, self-attention leads to high resiliency to noise (irrelevant features) compared to other benchmark models. Finally, we highlight the importance and separate contributions of both bottom-up and top-down attention in the learning process. Together, these results demonstrate the broader validity of the AWRL framework in complex task scenarios, and illustrate the benefits of a deeper integration between neuroscience-derived models and RL for decision making in machine learning.

Metacognitive resources for adaptive learning⋆.

Biological organisms display remarkably flexible behaviours. This is an area of active investigation, in particular in the fields of artificial intelligence, computational and cognitive neuroscience. While inductive biases and broader cognitive functions are undoubtedly important, the ability to monitor and evaluate one's performance or oneself -- metacognition -- strikes as a powerful resource for efficient learning. Often measured as decision confidence in neuroscience and psychology experiments, metacognition appears to reflect a broad range of abstraction levels and downstream behavioural effects. Within this context, the formal investigation of how metacognition interacts with learning processes is a recent endeavour. Of special interest are the neural and computational underpinnings of confidence and reinforcement learning modules. This review discusses a general hierarchy of confidence functions and their neuro-computational relevance for adaptive behaviours. It then introduces novel ways to study the formation and use of meta-representations and nonconscious mental representations related to learning and confidence, and concludes with a discussion on outstanding questions and wider perspectives.

Multiple time measurements of multidimensional psychiatric states from immediately before the COVID-19 pandemic to one year later: a longitudinal online survey of the Japanese population.

The coronavirus disease 2019 (COVID-19) pandemic has profoundly affected the mental health of both infected and uninfected people. Although most psychiatric disorders have highly overlapping genetic and pathogenic backgrounds, most studies investigating the impact of the pandemic have examined only single psychiatric disorders. It is necessary to examine longitudinal trajectories of factors that modulate psychiatric states across multiple dimensions. About 2274 Japanese citizens participated in online surveys presented in December 2019 (before the pandemic), August 2020, Dec 2020, and April 2021. These surveys included nine questionnaires on psychiatric symptoms, such as depression and anxiety. Multidimensional psychiatric time-series data were then decomposed into four principal components. We used generalized linear models to identify modulating factors for the effects of the pandemic on these components. The four principal components can be interpreted as a general psychiatric burden, social withdrawal, alcohol-related problems, and depression/anxiety. Principal components associated with general psychiatric burden and depression/anxiety peaked during the initial phase of the pandemic. They were further exacerbated by the economic burden the pandemic imposed. In contrast, principal components associated with social withdrawal showed a delayed peak, with human relationships as an important risk modulating factor. In addition, being female was a risk factor shared across all components. Our results show that COVID-19 has imposed a large and varied burden on the Japanese population since the commencement of the pandemic. Although components related to the general psychiatric burden remained elevated, peak intensities differed between components related to depression/anxiety and those related to social withdrawal. These results underline the importance of using flexible monitoring and mitigation strategies for mental problems, according to the phase of the pandemic.

From internal models toward metacognitive AI.

In several papers published in Biological Cybernetics in the 1980s and 1990s, Kawato and colleagues proposed computational models explaining how internal models are acquired in the cerebellum. These models were later supported by neurophysiological experiments using monkeys and neuroimaging experiments involving humans. These early studies influenced neuroscience from basic, sensory-motor control to higher cognitive functions. One of the most perplexing enigmas related to internal models is to understand the neural mechanisms that enable animals to learn large-dimensional problems with so few trials. Consciousness and metacognition-the ability to monitor one's own thoughts, may be part of the solution to this enigma. Based on literature reviews of the past 20 years, here we propose a computational neuroscience model of metacognition. The model comprises a modular hierarchical reinforcement-learning architecture of parallel and layered, generative-inverse model pairs. In the prefrontal cortex, a distributed executive network called the "cognitive reality monitoring network" (CRMN) orchestrates conscious involvement of generative-inverse model pairs in perception and action. Based on mismatches between computations by generative and inverse models, as well as reward prediction errors, CRMN computes a "responsibility signal" that gates selection and learning of pairs in perception, action, and reinforcement learning. A high responsibility signal is given to the pairs that best capture the external world, that are competent in movements (small mismatch), and that are capable of reinforcement learning (small reward-prediction error). CRMN selects pairs with higher responsibility signals as objects of metacognition, and consciousness is determined by the entropy of responsibility signals across all pairs. This model could lead to new-generation AI, which exhibits metacognition, consciousness, dimension reduction, selection of modules and corresponding representations, and learning from small samples. It may also lead to the development of a new scientific paradigm that enables the causal study of consciousness by combining CRMN and decoded neurofeedback.

Value signals guide abstraction during learning.

The human brain excels at constructing and using abstractions, such as rules, or concepts. Here, in two fMRI experiments, we demonstrate a mechanism of abstraction built upon the valuation of sensory features. Human volunteers learned novel association rules based on simple visual features. Reinforcement-learning algorithms revealed that, with learning, high-value abstract representations increasingly guided participant behaviour, resulting in better choices and higher subjective confidence. We also found that the brain area computing value signals - the ventromedial prefrontal cortex - prioritised and selected latent task elements during abstraction, both locally and through its connection to the visual cortex. Such a coding scheme predicts a causal role for valuation. Hence, in a second experiment, we used multivoxel neural reinforcement to test for the causality of feature valuation in the sensory cortex, as a mechanism of abstraction. Tagging the neural representation of a task feature with rewards evoked abstraction-based decisions. Together, these findings provide a novel interpretation of value as a goal-dependent, key factor in forging abstract representations.

The DecNef collection, fMRI data from closed-loop decoded neurofeedback experiments.

Decoded neurofeedback (DecNef) is a form of closed-loop functional magnetic resonance imaging (fMRI) combined with machine learning approaches, which holds some promises for clinical applications. Yet, currently only a few research groups have had the opportunity to run such experiments; furthermore, there is no existing public dataset for scientists to analyse and investigate some of the factors enabling the manipulation of brain dynamics. We release here the data from published DecNef studies, consisting of 5 separate fMRI datasets, each with multiple sessions recorded per participant. For each participant the data consists of a session that was used in the main experiment to train the machine learning decoder, and several (from 3 to 10) closed-loop fMRI neural reinforcement sessions. The large dataset, currently comprising more than 60 participants, will be useful to the fMRI community at large and to researchers trying to understand the mechanisms underlying non-invasive modulation of brain dynamics. Finally, the data collection size will increase over time as data from newly run DecNef studies will be added.

How do we generalize?

Humans and animals are able to generalize or transfer information from previous experience so that they can behave appropriately in novel situations. What mechanisms-computations, representations, and neural systems-give rise to this remarkable ability? The members of this Generative Adversarial Collaboration (GAC) come from a range of academic backgrounds but are all interested in uncovering the mechanisms of generalization. We started out this GAC with the aim of arbitrating between two alternative conceptual accounts: (1) generalization stems from integration of multiple experiences into summary representations that reflect generalized knowledge, and (2) generalization is computed on-the-fly using separately stored individual memories. Across the course of this collaboration, we found that-despite using different terminology and techniques, and although some of our specific papers may provide evidence one way or the other-we in fact largely agree that both of these broad accounts (as well as several others) are likely valid. We believe that future research and theoretical synthesis across multiple lines of research is necessary to help determine the degree to which different candidate generalization mechanisms may operate simultaneously, operate on different scales, or be employed under distinct conditions. Here, as the first step, we introduce some of these candidate mechanisms and we discuss the issues currently hindering better synthesis of generalization research. Finally, we introduce some of our own research questions that have arisen over the course of this GAC, that we believe would benefit from future collaborative efforts.

Conducting decoded neurofeedback studies.

Closed-loop neurofeedback has sparked great interest since its inception in the late 1960s. However, the field has historically faced various methodological challenges. Decoded fMRI neurofeedback may provide solutions to some of these problems. Notably, thanks to the recent advancements of machine learning approaches, it is now possible to target unconscious occurrences of specific multivoxel representations. In this tools of the trade paper, we discuss how to implement these interventions in rigorous double-blind placebo-controlled experiments. We aim to provide a step-by-step guide to address some of the most common methodological and analytical considerations. We also discuss tools that can be used to facilitate the implementation of new experiments. We hope that this will encourage more researchers to try out this powerful new intervention method.

Brain network dynamics fingerprints are resilient to data heterogeneity.

CONTEXT: Large multi-site neuroimaging datasets have significantly advanced our quest to understand brain-behavior relationships and to develop biomarkers of psychiatric and neurodegenerative disorders. Yet, such data collections come at a cost, as the inevitable differences across samples may lead to biased or erroneous conclusions. OBJECTIVE: We aim to validate the estimation of individual brain network dynamics fingerprints and appraise sources of variability in large resting-state functional magnetic resonance imaging (rs-fMRI) datasets by providing a novel point of view based on data-driven dynamical models. APPROACH: Previous work has investigated this critical issue in terms of effects on static measures, such as functional connectivity and brain parcellations. Here, we utilize dynamical models (Hidden Markov models - HMM) to examine how diverse scanning factors in multi-site fMRI recordings affect our ability to infer the brain's spatiotemporal wandering between large-scale networks of activity. Specifically, we leverage a stable HMM trained on the Human Connectome Project (homogeneous) dataset, which we then apply to an heterogeneous dataset of traveling subjects scanned under a multitude of conditions. MAIN RESULTS: Building upon this premise, we first replicate previous work on the emergence of non-random sequences of brain states. We next highlight how these time-varying brain activity patterns are robust subject-specific fingerprints. Finally, we suggest these fingerprints may be used to assess which scanning factors induce high variability in the data. SIGNIFICANCE: These results demonstrate that we can i) use large scale dataset to train models that can be then used to interrogate subject-specific data, ii) recover the unique trajectories of brain activity changes in each individual, but also iii) urge caution as our ability to infer such patterns is affected by how, where and when we do so.

Unconscious reinforcement learning of hidden brain states supported by confidence.

Can humans be trained to make strategic use of latent representations in their own brains? We investigate how human subjects can derive reward-maximizing choices from intrinsic high-dimensional information represented stochastically in neural activity. Reward contingencies are defined in real-time by fMRI multivoxel patterns; optimal action policies thereby depend on multidimensional brain activity taking place below the threshold of consciousness, by design. We find that subjects can solve the task within two hundred trials and errors, as their reinforcement learning processes interact with metacognitive functions (quantified as the meaningfulness of their decision confidence). Computational modelling and multivariate analyses identify a frontostriatal neural mechanism by which the brain may untangle the 'curse of dimensionality': synchronization of confidence representations in prefrontal cortex with reward prediction errors in basal ganglia support exploration of latent task representations. These results may provide an alternative starting point for future investigations into unconscious learning and functions of metacognition.

The Arctic/Swedish APP mutation alters the impact of chronic stress on cognition in mice.

Chronic stress is a major risk factor for developing Alzheimer's disease (AD) and promotes the processing of amyloid precursor protein (APP) to ß-amyloid (Aß). However, the precise relationship of stress and disease-typical cognitive decline is presently not well understood. The aim of this study was to investigate how early life stress may affect cognition in adult mice with and without soluble Aß pathology typical for the early stages of the disease. We focussed on sustained attention and response control, aspects of cognition mediated by the prefrontal cortex that are consistently impaired both in early AD and after chronic stress exposure. Young wild-type mice as well as transgenic arcAß mice overexpressing the hAPParc/swe transgene were exposed to a chronic unpredictable stress paradigm (age 3-8 weeks). At 15 weeks, these mice were tested on the 5-choice serial reaction time task, a test of sustained attention and executive control. We found that, expectedly, chronic stress increased impulsive choices and impaired sustained attention in wild-type mice. However, the same treatment reduced impulsivity and did not interfere with sustained attention in arcAß mice. These findings suggest an unexpected interaction between chronic stress and Aß whereby Aß-pathology caused by the hAPParc/swe mutation prevented and/or reversed stress-induced cognitive changes through mechanisms that deserve further investigation. They also indicate that Aß, in modest amounts, may have a beneficial role for cognitive stability, for example by protecting neural networks from the impact of further physiological or behavioural stress.

The neural and cognitive architecture for learning from a small sample.

Artificial intelligence algorithms are capable of fantastic exploits, yet they are still grossly inefficient compared with the brain's ability to learn from few exemplars or solve problems that have not been explicitly defined. What is the secret that the evolution of human intelligence has unlocked? Generalization is one answer, but there is more to it. The brain does not directly solve difficult problems, it is able to recast them into new and more tractable problems. Here, we propose a model whereby higher cognitive functions profoundly interact with reinforcement learning to drastically reduce the degrees of freedom of the search space, simplifying complex problems, and fostering more efficient learning.

Toward a comprehensive understanding of the neural mechanisms of decoded neurofeedback.

Real-time functional magnetic resonance imaging (fMRI) neurofeedback is an experimental framework in which fMRI signals are presented to participants in a real-time manner to change their behaviors. Changes in behaviors after real-time fMRI neurofeedback are postulated to be caused by neural plasticity driven by the induction of specific targeted activities at the neuronal level (targeted neural plasticity model). However, some research groups argued that behavioral changes in conventional real-time fMRI neurofeedback studies are explained by alternative accounts, including the placebo effect and physiological artifacts. Recently, decoded neurofeedback (DecNef) has been developed as a result of adapting new technological advancements, including implicit neurofeedback and fMRI multivariate analyses. DecNef provides strong evidence for the targeted neural plasticity model while refuting the abovementioned alternative accounts. In this review, we first discuss how DecNef refutes the alternative accounts. Second, we propose a model that shows how targeted neural plasticity occurs at the neuronal level during DecNef training. Finally, we discuss computational and empirical evidence that supports the model. Clarification of the neural mechanisms of DecNef would lead to the development of more advanced fMRI neurofeedback methods that may serve as powerful tools for both basic and clinical research.

Towards an unconscious neural reinforcement intervention for common fears.

Can "hardwired" physiological fear responses (e.g., for spiders and snakes) be reprogramed unconsciously in the human brain? Currently, exposure therapy is among the most effective treatments for anxiety disorders, but this intervention is subjectively aversive to patients, causing many to drop out of treatment prematurely. Here we introduce a method to bypass the subjective unpleasantness in conscious exposure, by directly pairing monetary reward with unconscious occurrences of decoded representations of naturally feared animals in the brain. To decode physiological fear representations without triggering excessively aversive reactions, we capitalize on recent advancements in functional magnetic resonance imaging decoding techniques, and use a method called hyperalignment to infer the relevant representations of feared animals for a designated participant based on data from other "surrogate" participants. In this way, the procedure completely bypasses the need for a conscious encounter with feared animals. We demonstrate that our method can lead to reliable reductions in physiological fear responses, as measured by skin conductance as well as amygdala hemodynamic activity. Not only do these results raise the intriguing possibility that naturally occurring fear responses can be "reprogrammed" outside of conscious awareness, importantly, they also create the rare opportunity to rigorously test a psychological intervention of this nature in a double-blind, placebo-controlled fashion. This may pave the way for a new approach combining the appealing rationale and proven efficacy of conventional psychotherapy with the rigor and leverage of clinical neuroscience.

[Modulation of Metacognition with Decoded Neurofeedback].

Humans often assess their confidence in their own perception, e.g., feeling "confident" or "certain" of having seen a friend, or feeling "uncertain" about whether the phone rang. The neural mechanism underlying the metacognitive function that reflects subjective perception still remains under debate. We have previously used decoded neurofeedback (DecNef) to demonstrate that manipulating the multivoxel activation patterns in the frontoparietal network modulates perceptual confidence without affecting perceptual performance. The results provided clear evidence for a dissociation between perceptual confidence and performance and suggested a distinct role of the frontoparietal network in metacognition.

Decoded fMRI neurofeedback can induce bidirectional confidence changes within single participants.

Neurofeedback studies using real-time functional magnetic resonance imaging (rt-fMRI) have recently incorporated the multi-voxel pattern decoding approach, allowing for fMRI to serve as a tool to manipulate fine-grained neural activity embedded in voxel patterns. Because of its tremendous potential for clinical applications, certain questions regarding decoded neurofeedback (DecNef) must be addressed. Specifically, can the same participants learn to induce neural patterns in opposite directions in different sessions? If so, how does previous learning affect subsequent induction effectiveness? These questions are critical because neurofeedback effects can last for months, but the short- to mid-term dynamics of such effects are unknown. Here we employed a within-subjects design, where participants underwent two DecNef training sessions to induce behavioural changes of opposing directionality (up or down regulation of perceptual confidence in a visual discrimination task), with the order of training counterbalanced across participants. Behavioral results indicated that the manipulation was strongly influenced by the order and the directionality of neurofeedback training. We applied nonlinear mathematical modeling to parametrize four main consequences of DecNef: main effect of change in confidence, strength of down-regulation of confidence relative to up-regulation, maintenance of learning effects, and anterograde learning interference. Modeling results revealed that DecNef successfully induced bidirectional confidence changes in different sessions within single participants. Furthermore, the effect of up- compared to down-regulation was more prominent, and confidence changes (regardless of the direction) were largely preserved even after a week-long interval. Lastly, the effect of the second session was markedly diminished as compared to the effect of the first session, indicating strong anterograde learning interference. These results are interpreted in the framework of reinforcement learning and provide important implications for its application to basic neuroscience, to occupational and sports training, and to therapy.

Multivoxel neurofeedback selectively modulates confidence without changing perceptual performance.

A central controversy in metacognition studies concerns whether subjective confidence directly reflects the reliability of perceptual or cognitive processes, as suggested by normative models based on the assumption that neural computations are generally optimal. This view enjoys popularity in the computational and animal literatures, but it has also been suggested that confidence may depend on a late-stage estimation dissociable from perceptual processes. Yet, at least in humans, experimental tools have lacked the power to resolve these issues convincingly. Here, we overcome this difficulty by using the recently developed method of decoded neurofeedback (DecNef) to systematically manipulate multivoxel correlates of confidence in a frontoparietal network. Here we report that bi-directional changes in confidence do not affect perceptual accuracy. Further psychophysical analyses rule out accounts based on simple shifts in reporting strategy. Our results provide clear neuroscientific evidence for the systematic dissociation between confidence and perceptual performance, and thereby challenge current theoretical thinking.