Adaptive Control Group

The lab works with two species: humans and (transgenic) mice. Despite obvious differences in neuroanatomy and cognitive abilities, these two models are very complementary to study the neurocomputational bases of adaptive control.

• On the one hand, the flexibility of human cognition allows investigating multiple cognitive mechanisms (and their interaction) in single individuals and neuroimaging (MEG,fMRI) techniques allows assessing the impact of cognitive tasks, disorders and treatments on the whole-brain with relative ease.

• On the other hand, transgenic lines, viral vectors and invasive techniques permitted by the mouse model allow measuring (calcium imaging, biosensors, Neuropixels) and manipulating (optogenetics) the activity of genetically-defined neurons, thereby enabling us to test highly constrained hypotheses regarding the computations of small neural circuits.

Computational modeling plays a pivotal role to design behavioral tasks able to capture the analogous (or even better: homologous) processes in both species and interpret the data within a unified framework.

Given our interest in neuromodulatory systems, observing similar effects after similar pharmacological manipulations represents another crucial validation step and a real perspective for the translation of system neuroscience into psychological and clinical insights.

Monoaminergic nuclei irrigate densely the basal ganglia and the cortex from their highly conserved position in brainstem circuits supporting homeostasis and reproduction. These peculiar properties allow them to regulate large-scale brain dynamics in a coordinated fashion through signalling at multiple receptors with different postsynaptic effects. Thus, unraveling the information encoded by dopamine (DA) and serotonin (5-HT) neurons are key to connecting computational theories and neurobiological phenomena.

In the case of DA, this idea is no longer a hypothesis since the activity of DA neurons was found to encode reward prediction errors and to provide a coherent link between reinforcement-learning and decision-making processes in various species [7–9]. This discovery has led computational neuroscientists to suggest complementary roles for 5-HT.

One of our key hypothesis is that 5-HT neurons of the dorsal raphe nucleus (DRN) broadcast controllability prediction errors signals and that this signal triggers changes in multiple neural circuits subserving prediction and action selection mechanisms, with important consequences for our understanding of depressive disorders and antidepressant treatments, amongst others. As recent studies indicate that DRN 5HT neurons might be fractionated in subsystems, we are also interested in the interplay of controllability signaling with other information conveyed by these neurons, such as locomotion, reinforcement and sensory prediction errors.

Our group is particularly interested in the adaptation to (perceived) fluctuations of controllability, that is, the degree to which the environment provides agents with opportunities to alter the course of events and reach specific states through action. Indeed, several lines of research, including ours, suggests that (perceived) controllability and related causal constructs such as agency, self-efficacy or empowerment may play important roles in the regulation of affective control and in the arbitration between different learning and decision-making systems, with clinically-relevant consequences for cognitive and brain function.

Rooted in causal inference principles and information-theory, the models we develop propose that agents willing to estimate the controllability of their environment can do so by comparing the predictions generated by two internal models of this environment, termed “spectator” and “actor”, which represent respectively state-state and state-action-state transition probabilities. Thus, one of their core variable is the controllability prediction error, which determines how internal estimates of controllability should be updated to account for ongoing observations and predict future ones.

Besides our focus in the neuromodulatory pathways broadcasting this second-order prediction error to the forebrain, we are also interested in the cortical and subcortical mechanisms underlying actor and spectator representations.

A fundamental goal of neuroscience is to understand why organisms perform certain actions in a given situation. Current theories suggest that (at least) two distinct decision mechanisms compete and interact. Indeed, while different schools of thought emphasize different features, most agree that mammalian nervous systems may act based on fast "reactive" heuristics requiring little efforts and cognitive resources or based on a slower "proactive" process integrating finer contextual information and anticipating the consequences of alternative action plans.

Computational models help us to articulate controllability estimation with other prominent adaptive mechanisms guiding action selection, such as reward-maximization (i.e., reinforcement-learning) and uncertainty-minimization (i.e., predictive coding), and to draw specific hypotheses regarding its influence on mammalian behavior and its neurobiological underpinnings.

Ours and others works suggest that perceived controllability may regulate the cognitive resources invested to solve adaptive problems by contributing to the arbitration between different decision-making systems associated with: (i) proactive versus reactive, (ii) goal-directed versus habitual or (iii) instrumental versus Pavlovian actions. Yet, the cognitive and neural mechanisms through which this regulation occur remain elusive. Our hope is that elucidating this question will pave the way to a better understanding of neuropsychiatric disorders associated with imbalances in controllability estimation, as well as better diagnosis and more personalized treatment.