Publications

Optimal inference with suboptimal models: Addiction and active Bayesian inference

When casting behaviour as active (Bayesian) inference, optimal inference is defined with respect to an agent's beliefs – based on its generative model of the world. This contrasts with normative accounts of choice behaviour, in which optimal actions are considered in relation to the true structure of the environment – as opposed to the agent's beliefs about worldly states (or the task). This distinction shifts an understanding of suboptimal or pathological behaviour away from aberrant inference as such, to understanding the prior beliefs of a subject that cause them to behave less 'optimally' than our prior beliefs suggest they should behave. Put simply, suboptimal or pathological behaviour does not speak against understanding behaviour in terms of (Bayes optimal) inference, but rather calls for a more refined understanding of the subject's generative model upon which their (optimal) Bayesian inference is based. Here, we discuss this fundamental distinction and its implications for understanding optimality, bounded rationality and pathological (choice) behaviour. We illustrate our argument using addictive choice behaviour in a recently described 'limited offer' task. Our simulations of pathological choices and addictive behaviour also generate some clear hypotheses, which we hope to pursue in ongoing empirical work.

Estimating anisotropy directly via neural timeseries

Abstract An isotropic dynamical system is one that looks the same in every direction, i.e., if we imagine standing somewhere within an isotropic system, we would not be able to differentiate between different lines of sight. Conversely, anisotropy is a measure of the extent to which a system deviates from perfect isotropy, with larger values indicating greater discrepancies between the structure of the system along its axes. Here, we derive the form of a generalised scalable (mechanically similar) discretized field theoretic Lagrangian that allows for levels of anisotropy to be directly estimated via timeseries of arbitrary dimensionality. We generate synthetic data for both isotropic and anisotropic systems and, by using Bayesian model inversion and reduction, show that we can discriminate between the two datasets – thereby demonstrating proof of principle. We then apply this methodology to murine calcium imaging data collected in rest and task states, showing that anisotropy can be estimated directly from different brain states and cortical regions in an empirical in vivo biological setting. We hope that this theoretical foundation, together with the methodology and publicly available MATLAB code, will provide an accessible way for researchers to obtain new insight into the structural organization of neural systems in terms of how scalable neural regions grow – both ontogenetically during the development of an individual organism, as well as phylogenetically across species.

Migraine as an allostatic reset triggered by unresolved interoceptive prediction errors

Until now, a satisfying account of the cause and purpose of migraine has remained elusive. We explain migraine within the frameworks of allostasis (the situationally-flexible, forward-looking equivalent of homeostasis) and active inference (interacting with the environment via internally-generated predictions). Due to its multimodality, and long timescales between cause and effect, allostasis is inherently prone to catastrophic error, which might be impossible to correct once fully manifest, an early indicator which is elevated prediction error (discrepancy between prediction and sensory input) associated with internal sensations (interoception). Errors can usually be resolved in a targeted manner by action (correcting the physiological state) or perception (updating predictions in light of sensory input); persistent errors are amplified broadly and multimodally, to prioritise their resolution (the migraine premonitory phase); finally, if still unresolved, progressive amplification renders further changes to internal or external sensory inputs intolerably intense, enforcing physiological stability, and facilitating accurate allostatic prediction updating. As such, migraine is an effective 'failsafe' for allostasis, however it has potential to become excessively triggered, therefore maladaptive.

Active Inference Tree Search in Large POMDPs

The ability to plan ahead efficiently is key for both living organisms and artificial systems. Model-based planning and prospection are widely studied in cognitive neuroscience and artificial intelligence (AI), but from different perspectives--and with different desiderata in mind (biological realism versus scalability) that are difficult to reconcile. Here, we introduce a novel method to plan in POMDPs--Active Inference Tree Search (AcT)--that combines the normative character and biological realism of a leading planning theory in neuroscience (Active Inference) and the scalability of tree search methods in AI. This unification enhances both approaches. On the one hand, tree searches enable the biologically grounded, first principle method of active inference to be applied to large-scale problems. On the other hand, active inference provides a principled solution to the exploration-exploitation dilemma, which is often addressed heuristically in tree search methods. Our simulations show that AcT successfully navigates binary trees that are challenging for sampling-based methods, problems that require adaptive exploration, and the large POMDP problem 'RockSample'--in which AcT reproduces state-of-the-art POMDP solutions. Furthermore, we illustrate how AcT can be used to simulate neurophysiological responses (e.g., in the hippocampus and prefrontal cortex) of humans and other animals that solve large planning problems. These numerical analyses show that Active Tree Search is a principled realisation of neuroscientific and AI planning theories, which offer both biological realism and scalability.

Dynamic causal modeling of evoked responses in EEG and MEG

Neuronally plausible, generative or forward models are essential for understanding how event-related fields (ERFs) and potentials (ERPs) are generated. In this paper, we present a new approach to modeling event-related responses measured with EEG or MEG. This approach uses a biologically informed model to make inferences about the underlying neuronal networks generating responses. The approach can be regarded as a neurobiologically constrained source reconstruction scheme, in which the parameters of the reconstruction have an explicit neuronal interpretation. Specifically, these parameters encode, among other things, the coupling among sources and how that coupling depends upon stimulus attributes or experimental context. The basic idea is to supplement conventional electromagnetic forward models, of how sources are expressed in measurement space, with a model of how source activity is generated by neuronal dynamics. A single inversion of this extended forward model enables inference about both the spatial deployment of sources and the underlying neuronal architecture generating them. Critically, this inference covers long-range connections among well-defined neuronal subpopulations. In a previous paper, we simulated ERPs using a hierarchical neural-mass model that embodied bottom-up, top-down and lateral connections among remote regions. In this paper, we describe a Bayesian procedure to estimate the parameters of this model using empirical data. We demonstrate this procedure by characterizing the role of changes in cortico-cortical coupling, in the genesis of ERPs. In the first experiment, ERPs recorded during the perception of faces and houses were modeled as distinct cortical sources in the ventral visual pathway. Category-selectivity, as indexed by the face-selective N170, could be explained by category-specific differences in forward connections from sensory to higher areas in the ventral stream. We were able to quantify and make inferences about these effects using conditional estimates of connectivity. This allowed us to identify where, in the processing stream, category-selectivity emerged. In the second experiment, we used an auditory oddball paradigm to show that the mismatch negativity can be explained by changes in connectivity. Specifically, using Bayesian model selection, we assessed changes in backward connections, above and beyond changes in forward connections. In accord with theoretical predictions, there was strong evidence for learning-related changes in both forward and backward coupling. These examples show that category- or context-specific coupling among cortical regions can be assessed explicitly, within a mechanistic, biologically motivated inference framework.

Making sense of neuroimaging data

Review of: "A brief introduction to the COM-B Model of behaviour and the PRIME Theory of motivation"

Author Correction: Cerebellar state estimation enables resilient coupling across behavioural domains

A follow up report validating long term predictions of the COVID-19 epidemic in the UK using a dynamic causal model

This paper asks whether Dynamic Causal modelling (DCM) can predict the long-term clinical impact of the COVID-19 epidemic. DCMs are designed to continually assimilate data and modify model parameters, such as transmissibility of the virus, changes in social distancing and vaccine coverage-to accommodate changes in population dynamics and virus behavior. But as a novel way to model epidemics do they produce valid predictions? We presented DCM predictions 12 months ago, which suggested an increase in viral transmission was accompanied by a reduction in pathogenicity. These changes provided plausible reasons why the model underestimated deaths, hospital admissions and acute-post COVID-19 syndrome by 20%. A further 12-month validation exercise could help to assess how useful such predictions are.

Waves of prediction

Predictive processing (e.g., predictive coding) is a predominant paradigm in cognitive neuroscience. This Primer considers the various levels of commitment neuroscientists have to the neuronal process theories that accompany the principles of predictive processing. Specifically, it reviews and contextualises a recent PLOS Biology study of alpha oscillations and travelling waves. We will see that alpha oscillations emerge naturally under the computational architectures implied by predictive coding-and may tell us something profound about recurrent message passing in brain hierarchies. Specifically, the bidirectional nature of forward and backward waves speaks to opportunities to understand attention and how it nuances bottom-up and top-down influences.

Role of the human rostral supplementary motor area and the basal ganglia in motor sequence control: Investigations with H-2 O-15 PET (vol 79, pg 1070, 1998)

Variational free energy, individual fitness, and population dynamics under acute stress

Self-Organization and Compositionality in Cognitive Brains

Bayesian inference, predictive coding and delusions

This paper considers psychotic symptoms in terms of false inferences or beliefs. It is based on the notion that the brain is an organ of inference that actively constructs hypotheses to explain or predict its sensations. This perspective provides a normative (Bayes optimal) account of action and perception that emphasises probabilistic representations; in particular, the confidence or precision of beliefs about the world. We consider sensory attenuation deficits, catatonia and delusions as various expressions of the same core pathology: namely, an aberrant encoding of precision in a predictive coding hierarchy. In predictive coding, precision is thought to be encoded by the postsynaptic gain of neurons reporting prediction error. This suggests that both pervasive trait abnormalities and florid failures of inference in the psychotic state can be linked to factors controlling postsynaptic gain-such as NMDA receptor function and (dopaminergic) neuromodulation. We illustrate these points using a biologically plausible simulation of attribution of agency-showing how a reduction in the precision of prior beliefs, relative to sensory evidence, can lead to false inference.

Modelling Trial-by-Trial Changes in the Mismatch Negativity

The mismatch negativity (MMN) is a differential brain response to violations of learned regularities. It has been used to demonstrate that the brain learns the statistical structure of its environment and predicts future sensory inputs. However, the algorithmic nature of these computations and the underlying neurobiological implementation remain controversial. This article introduces a mathematical framework with which competing ideas about the computational quantities indexed by MMN responses can be formalized and tested against single-trial EEG data. This framework was applied to five major theories of the MMN, comparing their ability to explain trial-by-trial changes in MMN amplitude. Three of these theories (predictive coding, model adjustment, and novelty detection) were formalized by linking the MMN to different manifestations of the same computational mechanism: approximate Bayesian inference according to the free-energy principle. We thereby propose a unifying view on three distinct theories of the MMN. The relative plausibility of each theory was assessed against empirical single-trial MMN amplitudes acquired from eight healthy volunteers in a roving oddball experiment. Models based on the free-energy principle provided more plausible explanations of trial-by-trial changes in MMN amplitude than models representing the two more traditional theories (change detection and adaptation). Our results suggest that the MMN reflects approximate Bayesian learning of sensory regularities, and that the MMN-generating process adjusts a probabilistic model of the environment according to prediction errors.

Reclaiming saliency: Rhythmic precision-modulated action and perception

Computational models of visual attention in artificial intelligence and robotics have been inspired by the concept of a saliency map. These models account for the mutual information between the (current) visual information and its estimated causes. However, they fail to consider the circular causality between perception and action. In other words, they do not consider where to sample next, given current beliefs. Here, we reclaim salience as an active inference process that relies on two basic principles: uncertainty minimization and rhythmic scheduling. For this, we make a distinction between attention and salience. Briefly, we associate attention with precision control, i.e., the confidence with which beliefs can be updated given sampled sensory data, and salience with uncertainty minimization that underwrites the selection of future sensory data. Using this, we propose a new account of attention based on rhythmic precision-modulation and discuss its potential in robotics, providing numerical experiments that showcase its advantages for state and noise estimation, system identification and action selection for informative path planning.

The labile brain. I. Neuronal transients and nonlinear coupling

In this, the first of three papers, the nature of, and motivation for, neuronal transients is described in relation to characterizing brain dynamics. This paper deals with some basic aspects of neuronal dynamics, interactions, coupling and implicit neuronal codes. The second paper develops neuronal transients and nonlinear coupling in the context of dynamic instability and complexity, and suggests that instability or lability is necessary for adaptive self–organization. The final paper addresses the role of neuronal transients through information theory and the emergence of spatio–temporal receptive fields and functional specialization. By considering the brain as an ensemble of connected dynamic systems one can show that a sufficient description of neuronal dynamics comprises neuronal activity at a particular time and its recent history. This history constitutes a neuronal transient. As such, transients represent a fundamental metric of neuronal interactions and, implicitly, a code employed in the functional integration of brain systems. The nature of transients, expressed conjointly in distinct neuronal populations, reflects the underlying coupling among populations. This coupling may be synchronous (and possibly oscillatory) or asynchronous. A critical distinction between synchronous and asynchronous coupling is that the former is essentially linear and the latter is nonlinear. The nonlinear nature of asynchronous coupling enables the rich, context–sensitive interactions that characterize real brain dynamics, suggesting that it plays a role in functional integration that may be as important as synchronous interactions. The distinction between linear and nonlinear coupling has fundamental implications for the analysis and characterization of neuronal interactions, most of which are predicated on linear (synchronous) coupling (e.g. crosscorrelograms and coherence). Using neuromagnetic data it is shown that nonlinear (asynchronous) coupling is, in fact, more abundant and can be more significant than synchronous coupling.

Bilinear dynamical systems

In this paper, we propose the use of bilinear dynamical systems (BDS)s for model-based deconvolution of fMRI time-series. The importance of this work lies in being able to deconvolve haemodynamic time-series, in an informed way, to disclose the underlying neuronal activity. Being able to estimate neuronal responses in a particular brain region is fundamental for many models of functional integration and connectivity in the brain. BDSs comprise a stochastic bilinear neurodynamical model specified in discrete time, and a set of linear convolution kernels for the haemodynamics. We derive an expectation-maximization (EM) algorithm for parameter estimation, in which fMRI time-series are deconvolved in an E-step and model parameters are updated in an M-Step. We report preliminary results that focus on the assumed stochastic nature of the neurodynamic model and compare the method to Wiener deconvolution.