Structure learning enhances concept formation in synthetic Active Inference agents.

Humans display astonishing skill in learning about the environment in which they operate. They assimilate a rich set of affordances and interrelations among different elements in particular contexts, and form flexible abstractions (i.e., concepts) that can be generalised and leveraged with ease. To capture these abilities, we present a deep hierarchical Active Inference model of goal-directed behaviour, and the accompanying belief update schemes implied by maximising model evidence. Using simulations, we elucidate the potential mechanisms that underlie and influence concept learning in a spatial foraging task. We show that the representations formed-as a result of foraging-reflect environmental structure in a way that is enhanced and nuanced by Bayesian model reduction, a special case of structure learning that typifies learning in the absence of new evidence. Synthetic agents learn associations and form concepts about environmental context and configuration as a result of inferential, parametric learning, and structure learning processes-three processes that can produce a diversity of beliefs and belief structures. Furthermore, the ensuing representations reflect symmetries for environments with identical configurations.

Item-level analysis of mental health symptom trajectories during the COVID-19 pandemic in the UK: Associations with age, sex and pre-existing psychiatric conditions.

BACKGROUND: There is widespread concern regarding how the COVID-19 pandemic has affected mental health. Emerging meta-analyses suggest that the impact on anxiety/depression may have been transient, but much of the included literature has major methodological limitations. Addressing this topic rigorously requires longitudinal data of sufficient scope and scale, controlling for contextual variables, with baseline data immediately pre-pandemic. AIMS: To analyse self-report of symptom frequency from two largely UK-based longitudinal cohorts: Cohort 1 (N = 10,475, two time-points: winter pre-pandemic to UK first winter resurgence), and Cohort 2 (N = 10,391, two time-points, peak first wave to UK first winter resurgence). METHOD: Multinomial logistic regression applied at the item level identified sub-populations with greater probability of change in mental health symptoms. Permutation analyses characterised changes in symptom frequency distributions. Cross group differences in symptom stability were evaluated via entropy of response transitions. RESULTS: Anxiety was the most affected aspect of mental health. The profiles of change in mood symptoms was less favourable for females and older adults. Those with pre-existing psychiatric diagnoses showed substantially higher probability of very frequent symptoms pre-pandemic and elevated risk of transitioning to the highest levels of symptoms during the pandemic. Elevated mental health symptoms were evident across intra-COVID timepoints in Cohort 2. CONCLUSIONS: These findings suggest that mental health has been negatively affected by the pandemic, including in a sustained fashion beyond the first UK lockdown into the first winter resurgence. Women, and older adults, were more affected relative to their own baselines. Those with diagnoses of psychiatric conditions were more likely to experience transition to the highest levels of symptom frequency.

Contextual perception under active inference.

Human social interactions depend on the ability to resolve uncertainty about the mental states of others. The context in which social interactions take place is crucial for mental state attribution as sensory inputs may be perceived differently depending on the context. In this paper, we introduce a mental state attribution task where a target-face with either an ambiguous or an unambiguous emotion is embedded in different social contexts. The social context is determined by the emotions conveyed by other faces in the scene. This task involves mental state attribution to a target-face (either happy or sad) depending on the social context. Using active inference models, we provide a proof of concept that an agent's perception of sensory stimuli may be altered by social context. We show with simulations that context congruency and facial expression coherency improve behavioural performance in terms of decision times. Furthermore, we show through simulations that the abnormal viewing strategies employed by patients with schizophrenia may be due to (i) an imbalance between the precisions of local and global features in the scene and (ii) a failure to modulate the sensory precision to contextualise emotions.

Generative Models for Active Vision.

The active visual system comprises the visual cortices, cerebral attention networks, and oculomotor system. While fascinating in its own right, it is also an important model for sensorimotor networks in general. A prominent approach to studying this system is active inference-which assumes the brain makes use of an internal (generative) model to predict proprioceptive and visual input. This approach treats action as ensuring sensations conform to predictions (i.e., by moving the eyes) and posits that visual percepts are the consequence of updating predictions to conform to sensations. Under active inference, the challenge is to identify the form of the generative model that makes these predictions-and thus directs behavior. In this paper, we provide an overview of the generative models that the brain must employ to engage in active vision. This means specifying the processes that explain retinal cell activity and proprioceptive information from oculomotor muscle fibers. In addition to the mechanics of the eyes and retina, these processes include our choices about where to move our eyes. These decisions rest upon beliefs about salient locations, or the potential for information gain and belief-updating. A key theme of this paper is the relationship between "looking" and "seeing" under the brain's implicit generative model of the visual world.

Deep Active Inference and Scene Construction.

Adaptive agents must act in intrinsically uncertain environments with complex latent structure. Here, we elaborate a model of visual foraging-in a hierarchical context-wherein agents infer a higher-order visual pattern (a "scene") by sequentially sampling ambiguous cues. Inspired by previous models of scene construction-that cast perception and action as consequences of approximate Bayesian inference-we use active inference to simulate decisions of agents categorizing a scene in a hierarchically-structured setting. Under active inference, agents develop probabilistic beliefs about their environment, while actively sampling it to maximize the evidence for their internal generative model. This approximate evidence maximization (i.e., self-evidencing) comprises drives to both maximize rewards and resolve uncertainty about hidden states. This is realized via minimization of a free energy functional of posterior beliefs about both the world as well as the actions used to sample or perturb it, corresponding to perception and action, respectively. We show that active inference, in the context of hierarchical scene construction, gives rise to many empirical evidence accumulation phenomena, such as noise-sensitive reaction times and epistemic saccades. We explain these behaviors in terms of the principled drives that constitute the expected free energy, the key quantity for evaluating policies under active inference. In addition, we report novel behaviors exhibited by these active inference agents that furnish new predictions for research on evidence accumulation and perceptual decision-making. We discuss the implications of this hierarchical active inference scheme for tasks that require planned sequences of information-gathering actions to infer compositional latent structure (such as visual scene construction and sentence comprehension). This work sets the stage for future experiments to investigate active inference in relation to other formulations of evidence accumulation (e.g., drift-diffusion models) in tasks that require planning in uncertain environments with higher-order structure.

Introducing a Bayesian model of selective attention based on active inference.

Information gathering comprises actions whose (sensory) consequences resolve uncertainty (i.e., are salient). In other words, actions that solicit salient information cause the greatest shift in beliefs (i.e., information gain) about the causes of our sensations. However, not all information is relevant to the task at hand: this is especially the case in complex, naturalistic scenes. This paper introduces a formal model of selective attention based on active inference and contextual epistemic foraging. We consider a visual search task with a special emphasis on goal-directed and task-relevant exploration. In this scheme, attention modulates the expected fidelity (precision) of the mapping between observations and hidden states in a state-dependent or context-sensitive manner. This ensures task-irrelevant observations have little expected information gain, and so the agent - driven to reduce expected surprise (i.e., uncertainty) - does not actively seek them out. Instead, it selectively samples task-relevant observations, which inform (task-relevant) hidden states. We further show, through simulations, that the atypical exploratory behaviours in conditions such as autism and anxiety may be due to a failure to appropriately modulate sensory precision in a context-specific way.

Dynamic Causal Modelling of Active Vision.

In this paper, we draw from recent theoretical work on active perception, which suggests that the brain makes use of an internal (i.e., generative) model to make inferences about the causes of sensations. This view treats visual sensations as consequent on action (i.e., saccades) and implies that visual percepts must be actively constructed via a sequence of eye movements. Oculomotor control calls on a distributed set of brain sources that includes the dorsal and ventral frontoparietal (attention) networks. We argue that connections from the frontal eye fields to ventral parietal sources represent the mapping from "where", fixation location to information derived from "what" representations in the ventral visual stream. During scene construction, this mapping must be learned, putatively through changes in the effective connectivity of these synapses. Here, we test the hypothesis that the coupling between the dorsal frontal cortex and the right temporoparietal cortex is modulated during saccadic interrogation of a simple visual scene. Using dynamic causal modeling for magnetoencephalography with (male and female) human participants, we assess the evidence for changes in effective connectivity by comparing models that allow for this modulation with models that do not. We find strong evidence for modulation of connections between the two attention networks; namely, a disinhibition of the ventral network by its dorsal counterpart.SIGNIFICANCE STATEMENT This work draws from recent theoretical accounts of active vision and provides empirical evidence for changes in synaptic efficacy consistent with these computational models. In brief, we used magnetoencephalography in combination with eye-tracking to assess the neural correlates of a form of short-term memory during a dot cancellation task. Using dynamic causal modeling to quantify changes in effective connectivity, we found evidence that the coupling between the dorsal and ventral attention networks changed during the saccadic interrogation of a simple visual scene. Intuitively, this is consistent with the idea that these neuronal connections may encode beliefs about "what I would see if I looked there", and that this mapping is optimized as new data are obtained with each fixation.

Impulsivity and Active Inference.

This paper characterizes impulsive behavior using a patch-leaving paradigm and active inference-a framework for describing Bayes optimal behavior. This paradigm comprises different environments (patches) with limited resources that decline over time at different rates. The challenge is to decide when to leave the current patch for another to maximize reward. We chose this task because it offers an operational characterization of impulsive behavior, namely, maximizing proximal reward at the expense of future gain. We use a Markov decision process formulation of active inference to simulate behavioral and electrophysiological responses under different models and prior beliefs. Our main finding is that there are at least three distinct causes of impulsive behavior, which we demonstrate by manipulating three different components of the Markov decision process model. These components comprise (i) the depth of planning, (ii) the capacity to maintain and process information, and (iii) the perceived value of immediate (relative to delayed) rewards. We show how these manipulations change beliefs and subsequent choices through variational message passing. Furthermore, we appeal to the process theories associated with this message passing to simulate neuronal correlates. In future work, we will use this scheme to identify the prior beliefs that underlie different sorts of impulsive behavior-and ask whether different causes of impulsivity can be inferred from the electrophysiological correlates of choice behavior.

Human visual exploration reduces uncertainty about the sensed world.

In previous papers, we introduced a normative scheme for scene construction and epistemic (visual) searches based upon active inference. This scheme provides a principled account of how people decide where to look, when categorising a visual scene based on its contents. In this paper, we use active inference to explain the visual searches of normal human subjects; enabling us to answer some key questions about visual foraging and salience attribution. First, we asked whether there is any evidence for 'epistemic foraging'; i.e. exploration that resolves uncertainty about a scene. In brief, we used Bayesian model comparison to compare Markov decision process (MDP) models of scan-paths that did-and did not-contain the epistemic, uncertainty-resolving imperatives for action selection. In the course of this model comparison, we discovered that it was necessary to include non-epistemic (heuristic) policies to explain observed behaviour (e.g., a reading-like strategy that involved scanning from left to right). Despite this use of heuristic policies, model comparison showed that there is substantial evidence for epistemic foraging in the visual exploration of even simple scenes. Second, we compared MDP models that did-and did not-allow for changes in prior expectations over successive blocks of the visual search paradigm. We found that implicit prior beliefs about the speed and accuracy of visual searches changed systematically with experience. Finally, we characterised intersubject variability in terms of subject-specific prior beliefs. Specifically, we used canonical correlation analysis to see if there were any mixtures of prior expectations that could predict between-subject differences in performance; thereby establishing a quantitative link between different behavioural phenotypes and Bayesian belief updating. We demonstrated that better scene categorisation performance is consistently associated with lower reliance on heuristics; i.e., a greater use of a generative model of the scene to direct its exploration.

Scene Construction, Visual Foraging, and Active Inference.

This paper describes an active inference scheme for visual searches and the perceptual synthesis entailed by scene construction. Active inference assumes that perception and action minimize variational free energy, where actions are selected to minimize the free energy expected in the future. This assumption generalizes risk-sensitive control and expected utility theory to include epistemic value; namely, the value (or salience) of information inherent in resolving uncertainty about the causes of ambiguous cues or outcomes. Here, we apply active inference to saccadic searches of a visual scene. We consider the (difficult) problem of categorizing a scene, based on the spatial relationship among visual objects where, crucially, visual cues are sampled myopically through a sequence of saccadic eye movements. This means that evidence for competing hypotheses about the scene has to be accumulated sequentially, calling upon both prediction (planning) and postdiction (memory). Our aim is to highlight some simple but fundamental aspects of the requisite functional anatomy; namely, the link between approximate Bayesian inference under mean field assumptions and functional segregation in the visual cortex. This link rests upon the (neurobiologically plausible) process theory that accompanies the normative formulation of active inference for Markov decision processes. In future work, we hope to use this scheme to model empirical saccadic searches and identify the prior beliefs that underwrite intersubject variability in the way people forage for information in visual scenes (e.g., in schizophrenia).