Publications

The exact mechanisms whereby the cholinergic neurotransmitter system contributes to attentional processing remain poorly understood. Here, we applied computational modeling to psychophysical data (obtained from a spatial attention task) under a psychopharmacological challenge with the cholinesterase inhibitor galantamine (Reminyl). This allowed us to characterize the cholinergic modulation of selective attention formally, in terms of hierarchical Bayesian inference. In a placebo-controlled, within-subject, crossover design, 16 healthy human subjects performed a modified version of Posner's location-cueing task in which the proportion of validly and invalidly cued targets (percentage of cue validity, % CV) changed over time. Saccadic response speeds were used to estimate the parameters of a hierarchical Bayesian model to test whether cholinergic stimulation affected the trial-wise updating of probabilistic beliefs that underlie the allocation of attention or whether galantamine changed the mapping from those beliefs to subsequent eye movements. Behaviorally, galantamine led to a greater influence of probabilistic context (% CV) on response speed than placebo. Crucially, computational modeling suggested this effect was due to an increase in the rate of belief updating about cue validity (as opposed to the increased sensitivity of behavioral responses to those beliefs). We discuss these findings with respect to cholinergic effects on hierarchical cortical processing and in relation to the encoding of expected uncertainty or precision.

Electrophysiological studies on monkeys have been able to distinguish sensory and motor signals close in time by pseudo- randomly delaying the cue that instructs the movement from the stimulus that triggers the movement. We have used a similar experimental design in functional magnetic resonance imaging (fMRI), scanning subjects while they performed a visuomotor conditional task with instructed delays. One of four shapes was presented briefly. Two shapes instructed the subjects to flex the index finger; the other two shapes coded the flexion of the middle finger. The subjects were told to perform the movement after a tone. We have exploited a novel use of event-related fMRI. By systematically varying the interval between the visual and acoustic stimuli, it has been possible to estimate the significance of the evoked hæmodynamic response (EHR) to each of the stimuli, despite their temporal proximity in relation to the time constant of the EHR. Furthermore, by varying the phase between events and image acquisition, we have been able to achieve high temporal resolution while scanning the whole brain. We dissociated sensory and motor components of the sensorimotor transformations elicited by the task, and assessed sustained activity during the instructed delays. In calcarine and occipitotemporal cortex, the responses were exclusively associated with the visual instruction cues. In temporal auditory cortex and in primary motor cortex, they were exclusively associated with the auditory trigger stimulus. In ventral prefrontal cortex there were movement-related responses preceded by preparatory activity and by signal-related activity. Finally, responses associated with the instruction cue and with sustained activity during the delay period were observed in the dorsal premotor cortex and in the dorsal posterior parietal cortex. Where the association between a visual cue and the appropriate movement is arbitrary, the underlying visuomotor transformations are not achieved exclusively through frontoparietal interactions. Rather, these processes seem to rely on the ventral visual stream, the ventral prefrontal cortex and the anterior part of the dorsal premotor cortex.

This paper investigates the prospect of developing human-interpretable, explainable artificial intelligence (AI) systems based on active inference and the free energy principle. We first provide a brief overview of active inference, and in particular, of how it applies to the modeling of decision-making, introspection, as well as the generation of overt and covert actions. We then discuss how active inference can be leveraged to design explainable AI systems, namely, by allowing us to model core features of ``introspective'' processes and by generating useful, human-interpretable models of the processes involved in decision-making. We propose an architecture for explainable AI systems using active inference. This architecture foregrounds the role of an explicit hierarchical generative model, the operation of which enables the AI system to track and explain the factors that contribute to its own decisions, and whose structure is designed to be interpretable and auditable by human users. We outline how this architecture can integrate diverse sources of information to make informed decisions in an auditable manner, mimicking or reproducing aspects of human-like consciousness and introspection. Finally, we discuss the implications of our findings for future research in AI, and the potential ethical considerations of developing AI systems with (the appearance of) introspective capabilities.

Successful behaviour depends on the right balance between maximising reward and soliciting information about the world. Here, we show how different types of information-gain emerge when casting behaviour as surprise minimisation. We present two distinct mechanisms for goal-directed exploration that express separable profiles of active sampling to reduce uncertainty. 'Hidden state' exploration motivates agents to sample unambiguous observations to accurately infer the (hidden) state of the world. Conversely, 'model parameter' exploration, compels agents to sample outcomes associated with high uncertainty, if they are informative for their representation of the task structure. We illustrate the emergence of these types of information-gain, termed active inference and active learning, and show how these forms of exploration induce distinct patterns of 'Bayes-optimal' behaviour. Our findings provide a computational framework for understanding how distinct levels of uncertainty systematically affect the exploration-exploitation trade-off in decision-making.

Various methods have been proposed to characterize the functional connectivity between nodes in a network measured with different modalities (electrophysiology, functional magnetic resonance imaging etc.). Since different measures of functional connectivity yield different results for the same dataset, it is important to assess when and how they can be used. In this work, we provide a systematic framework for evaluating the performance of a large range of functional connectivity measures – based upon a comprehensive portfolio of models generating measurable responses. Specifically, we benchmarked 42 methods using 10,000 simulated datasets from 5 different types of generative models with different connectivity structures. Since all functional connectivity methods require the setting of some parameters (window size and number, model order etc.), we first optimized these parameters using performance criteria based upon (threshold free) ROC analysis. We then evaluated the performance of the methods on data simulated with different types of models. Finally, we assessed the performance of the methods against different levels of signal-to-noise ratios and network configurations. A MATLAB toolbox is provided to perform such analyses using other methods and simulated datasets.

We advance a novel formulation of cognitive control within the active inference framework. The theory proposes that cognitive control amounts to optimising a precision parameter, which acts as a control signal and balances the contributions of deliberative and habitual components of action selection. To illustrate the theory, we simulate a driving scenario in which the driver follows a well-known route, but encounters unexpected challenges. Our simulations show that a standard active inference model can form adaptive habits; i.e., can pass from deliberative to habitual control when the context is stable, but generally fails to revert to deliberative control, when the context changes. To address this failure of context-sensitivity, we introduce a novel type of hierarchical active inference, in which a lower level is responsible for behavioural control and the higher (or meta-cognitive) level observes the belief updating of the lower level below and is responsible for cognitive control. Crucially, the meta-cognitive level can both form habits and suspend them, by controlling the (precision) parameter that prioritizes deliberative choices at the behavioural level. Furthermore, we show that several processes linked to cognitive control — such as surprise detection, cognitive conflict monitoring, control signal regulation and specification, the simulation of future outcomes and the assessment of the costs of control and mental effort — stem coherently from the free energy minimization scheme that underpins active inference. Finally, we discuss the putative neurobiology of cognitive control by simulating brain dynamics in the mesolimbic and mesocortical pathways of the dopamine system, the dorsal anterior cingulate cortex and the locus coeruleus.

Accurate perceptual inference fundamentally depends upon accurate beliefs about the reliability of sensory data. In this paper, we describe a Bayes optimal and biologically plausible scheme that refines these beliefs through a gradient descent on variational free energy. To illustrate this, we simulate belief updating during visual foraging and show that changes in estimated sensory precision (i.e. confidence in visual data) are highly sensitive to prior beliefs about the contents of a visual scene. In brief, confident prior beliefs induce an increase in estimated precision when consistent with sensory evidence, but a decrease when they conflict. Prior beliefs held with low confidence are rapidly updated to posterior beliefs, determined by sensory data. These induce much smaller changes in beliefs about sensory precision. We argue that pathologies of scene construction may be due to abnormal priors, and show that these can induce a reduction in estimated sensory precision. Having previously associated this precision with cholinergic signalling, we note that several neurodegenerative conditions are associated with visual disturbances and cholinergic deficits; notably, the synucleinopathies. On relating the message passing in our model to the functional anatomy of the ventral visual stream, we find that simulated neuronal loss in temporal lobe regions induces confident, inaccurate, empirical prior beliefs at lower levels in the visual hierarchy. This provides a plausible, if speculative, computational mechanism for the loss of cholinergic signalling and the visual disturbances associated with temporal lobe Lewy body pathology. This may be seen as an illustration of the sorts of hypotheses that may be expressed within this computational framework.