The Lab are research interests span from broad theoretical questions concerning the neurocomputational underpinnings of decision-making, goal-oriented behaviour and habit formation to specific investigation into the alterations of these neurocomputational processes underlying psychiatric disorders. Each theme informs the others: our computational models of behaviour and neural dynamics shape, and are shaped by, our understanding of clinical phenomena. This layered approach bridges the gap between theory, mechanism, and clinical interpretation.
Theme 1: Computational Psychiatry
Computational Psychiatry is a rapidly growing field that applies formal models to understand the mechanisms underlying healthy and aberrant behaviour and cognitive processing. The field was formally outlined in the landmark paper ‘Computational Psychiatry’ (Montague et al. 2012), which proposed the integration of computational modelling with clinical research as a transformative approach to psychiatry.
At the heart of Computational Psychiatry lies the idea that psychiatric symptoms reflect disruptions in fundamental computational processes—such as learning, decision-making, inference, or perception—and that these processes can be rigorously described, simulated, and tested using mathematical models. A classic way to understand our work in this field is by relying on David Marr’s influential framework of three levels of analysis (Marr, 1982, Vision – A Computational Investigation into the Human Representation and Processing of Visual Information):
-
Computational level
At this level, we ask why the system is performing a given computation or controlling an observed behaviour. In psychiatric research, this often involves formalising behaviour in terms of Bayesian inference, which provides a normative framework for how agents should integrate prior beliefs with incoming evidence to make decisions under uncertainty. We use Bayesian models to characterise how individuals generate and update predictions about the world and how these processes may go awry in clinical populations. -
Algorithmic level
This level specifies what computations are carried out to solve a given problem. We primarily use reinforcement learning (RL) models—such as model-free and model-based RL—to capture how people learn from rewards and punishments over time and how this information is stored, recalled and used in forward planning. These models allow us to quantify key aspects of learning and decision-making, including learning rates, value estimation, and exploration strategies, and to examine how they vary across individuals and disorders. -
Implementational level
Finally, we investigate where and how these computations are realised in the brain. We use neural network models, effective connectivity analysis, and other network neuroscience tools to identify the neural architectures responsible for implementing those inference and learning processes described in the previous two levels. Combining fMRI, EEG, and behavioural data, we seek to uncover how deviations in neural dynamics contribute to psychiatric symptoms.
By linking these three levels of analysis, our approach aims to bridge the gap between theoretical models, empirical observations, and clinical relevance. Ultimately, Computational Psychiatry offers a principled way to move from symptom-based diagnosis toward mechanistic understanding and personalised treatment strategies.
Theme 2: Cortico-striatal dynamics and evolutionary conserved neurocomputational mechanisms for decision-making
Cortico-striatal circuits, which include different regions of the cortex, their partially segregated targets in the thalamus and striatum, and functionally separated (but similarly structured) sections of the basal ganglia, are central to the selection, maintenance, and modulation of actions, goals and attention foci. These circuits are among the most evolutionarily conserved structures in the vertebrate brain, with strong anatomical and functional analogues also found in invertebrate species (Fiore et al 2015). This cross-species conservation points to their fundamental role in adaptive behaviour. Our lab draws on this deep evolutionary perspective to investigate the relation between their dynamics under aminergic modulation and their associated core computations, aiming at identifying what dysfunctions contribute to neuropsychiatric disorders.
The classical view, dominant since the late 1990s thank to seminal work by computational researchers such as K Gurney , T J Prescott, P Redgrave or Michael J Frank, suggested that the basal ganglia resolve action selection through a mechanism of focused facilitation and surround inhibition or control, which has been often simplified in a binary Go/NoGo function. However, in the past 10–15 years, both empirical findings and network-level modelling have challenged this framework, revealing a more dynamic and complex relation between neural architecture and its functions.
Building on this revised understanding, Vincenzo Fiore (PI) has proposed an alternative theory of basal ganglia function. Rather than acting simply as a gate for Go or NoGo responses, the basal ganglia are thought to play a role in stabilising or disrupting cortico-thalamic dynamics. Through their position in recurrent cortico-basal ganglia-thalamo-cortical loops, these structures are ideally placed to shape the emergence or dissolution of circuit-wide attractor states, resulting in patterns of neural activity that underlie the sustained or unstable representation of policies, goals, or attentional sets.
Under this account, the basal ganglia are not only involved in action initiation but also in the maintenance of internally consistent behavioural states, and in their flexible reorganisation when environmental contingencies demand change. This theory has been used to explain observed alterations in:
-
Attentional and behavioural variability in ADHD (Hauser, Fiore et al., Trends in Neurosciences, 2016)
-
Motor switching and rigidity, including under low dopaminergic conditions (Fiore et al., Scientific Reports, 2016, Fiore et al. Eur J Neurosci. 2021)
-
Neurocomputational markers of compulsivity in addiction (Fiore et al., eNeuro, 2018)
-
Computational markers of compulsivity in eating disorders (Berner, Fiore et al. Transl Psychiatry. 2023)
This line of research aims to offer a unifying computational framework for cortico-striatal function, grounded in evolutionary neurobiology and validated across diverse clinical conditions.
Theme 3: Disorders of Compulsivity
Disorders of compulsivity (Voon et al. 2015) —including substance use disorders, behavioural addictions, eating disorders, and obsessive-compulsive disorder (OCD)—are marked by persistent, repetitive behaviours that continue despite negative consequences and a desire to stop. These conditions often share a bias towards impaired flexibility, which has been ascribed to a common tendency to over-rely on habitual responses at the expense of goal-directed control.
Our lab examines compulsivity from the perspective of cortico-striatal circuit dynamics, particularly focusing on how the stability, instability, and metastability of neural activity in these circuits can give rise to persistent behavioural states. Building on our previous work on basal ganglia dynamics, we hypothesise that aberrant stability within cortico-striatal loops is a hallmark of compulsive pathology. This excessive stability may prevent the flexible reconfiguration of internal states, making behaviours resistant to change even when environmental contingencies shift.
Importantly, this dysregulation may manifest in both:
-
Dorsal striatal circuits, contributing to the development of rigid motor routines and habitual responses
-
Ventral and associative circuits, leading to persistent goal fixations or intrusive, obsession-like patterns
The presence of aberrant stability in associative and ventral circuits also implies that, for instance, in substance use disorders a drug related goal can be triggered in the absence of a reinforced cue, in a mechanism that would explain how drug use (like other compulsive behaviours) can transcend simple stimulus-response mappings (e.g. see Fiore et al., eNeuro, 2018). Using ad hoc developed computational models based on Bayesian inference, we are currently testing these hypotheses across diagnostic groups, to identify transdiagnostic and disorder specific neurocomputational signatures of compulsivity (e.g. see Kato et al. 2022; Berner, Fiore et al 2023).
A second, complementary line of work investigates how goal-directed behaviour and forward planning are affected in compulsivity. In high-complexity environments, individuals with compulsive tendencies may adopt maladaptive, overly deterministic goal structures. This hypothesis is formalised in our model of bounded rationality and environmental complexity in addiction (Ognibene, Fiore, & Gu, Neural Networks, 2019), which suggests that compulsivity may reflect a cognitive trade-off between computational load and behavioural flexibility.