Computational Modeling

We start with the premise that most basic decisions we make (e.g., in the form of choices or effort allocation) can be traced back in the structure of macro-scale brain activity, as measured with modern neuroimaging apparatus. Typically, such responses involve many regions in the brain (from midbrain nuclei to basal ganglia, to limbic and prefrontal cortices), whose precise function in terms of motivational processes depends upon the context (e.g., the specific task the brain is solving). This context-dependency expresses itself through the (induced) specific plasticity of these brain networks, in parallel to phasic and tonic changes in neuromodulatory activity. In turn, this macro-scale reconfiguration of brain networks subtends learning and yield (mal-)adaptive behaviour.

In other words, it is very likely that goal-directed behaviour emerges from the very same interactions that shape the spatio-temporal dynamics of macro-scale brain networks.

This means that understanding the mechanics of motivational processes from the multimodal observation of brain activity (e.g., electrophysiology, fMRI...) and behavioural measurements (e.g., explicit choices, reaction times, autonomic arousal signals, grip force...) poses the exciting challenge of quantitatively relating information processing to brain effective connectivity.

On the one hand, the long-term objective of our research effort is to finesse the problem of predicting clinical and behavioural outcomes from basic psycho-physiological interactions. For example, can we predict the effect of a drug (e.g., a dopamine reuptake inhibitor) on a given patient suffering from a specific motivational deficit (e.g., depression), who has been profiled using dedicated neuroeconomics methods (e.g., an fMRI examination during a rewarded effort management task)? Developing quantitative approaches that can do this will require merging expert knowledge on neurobiology, cognitive psychology and neuroimaging data analysis, to mention but a few.

On the other hand, our short-term endeavour is to build models and propose methods that serve experimental purposes, based on a few quantitative theories that have the potential to capture the richness of neurophysiological and behavioural responses. A key notion here is that all models are embedded into a formal statistical data analysis framework. This is required to performing a quantitative interpretation of experimental data (parameter estimation and model comparison), as well as designing novel experimental studies.

In this context, our current research projects address the development of:

i. Biophysically realistic models of macro-scale brain dynamics.

These are inspired from statistical physics approaches based upon the notion of mean field, i.e. the idea that interactions among micro-scale ensembles of neurons can be captured by summary statistics (i.e., moments of the relevant distribution). They describe the spatio-temporal response of brain networks to the experimental manipulation. The inversion of such models given neuroimaging data (see iii.) can then be used to identify the structure of brain networks and their specific modulation by the experimental manipulation. A related challenge is to bridge the gap between micro-scale measurements performed invasively in monkeys (e.g., multi-site single unit recordings) and macro-scale signals acquired routinely in humans (e.g., whole-brain fMRI). Yet another change of scale relates to how brain networks coordinate their activity to produce overt behaviour (see this paper for an exemple in the context of DCM).

ii. Probabilistic models of learning and decision making.

These are based upon Bayesian decision theory, i.e. a probabilistic account of how information is processed and decisions or actions are emitted. The inversion of such models (see iii.) is then used to identify the subjective beliefs and preferences that interact with the experimental manipulation to generate observed behavioural data (see this paper for an exemple in the context of mentalizing). In addition, these computational models are used to inform the identification of brain networks (see i.) implementing cognitive processes such as learning and decision making, given neuroimaging data. A key question here is to understand the link between neurophysiological variables (e.g., neural activity and indices of network plasticity) and computational processes (e.g., belief updates and goal-directed actions).

iii. Statistical techniques embedding the above models for analyzing neuroimaging and behavioural data.

These are probabilistic inversion schemes that borrow from disciplines such as inverse problems and machine learning. If only, they are necessary to capture the inter-individual variability of neurophysiological and behavioural responses. More generally, they are essential to root a principled approach to model comparison and selection, given experimental data (see this paper for an exposition of the VBA toolbox). This is important to identify candidate psycho-physiological scenarios that have the ability to quantitatively explain concurrent neuroimaging and behavioural data.

An ambition here is to construct a modelling framework general enough to relate the various experimental studies conducted in the group with each other, without compromising predictive power. One of the main difficulties is thus to balance the complexity of the above models with the sophistication of the experimental design and data analysis procedures. This simply means that these three aspects of the research have to be conducted in parallel. This joint effort towards a quantitative psycho-physiological understanding of motivation is what we term 'computational neuroeconomics'.