In an effort to further our understanding of brain systems, our research focuses on deciphering the functions of neural circuits linking frontal cortex with basal ganglia in cognitive processes such as learning, memory, decision making, and cognitive control. Such an understanding can provide insights into how these functions are altered by neurological and psychiatric conditions known to involve neurochemical and/or structural changes to this circuitry.


 Linking across multiple levels of analysis in model-based cognitive neuroscience


Our lab develops computational methods linking multiple levels of analysis to guide experiments and interpret data, affording a principled understanding of the neural mechanisms underlying motivation, learning and cognitive function.  Through this “interactionist” approach

we have contributed to the burgeoning field of computational psychiatry and neurology which promises to clarify the neuroscience behind mental and point the way to novel rational drug and brain stimulation therapies.

Progress requires intensive and rigorous investigation across multiple levels as well as machine learning methods for appropriate clustering and classification of individuals. My work has focused on combining these multi-level studies to reciprocally inform the other, enriching the potential for understanding provided by any one level. We have developed and refined computational models of neural circuits linking frontal cortex with basal ganglia, and have tested model predictions using a combination of neuropsychological, pharmacological, genetic, brain stimulation, and neuroimaging techniques across species.

For theory development, we leverage a combinati

on of three distinct levels of computation.

* First, we characterize cortical and basal ganglia circuits with dynamical neural systems models,  simulating patterns of electrophysiology in distinct regions and cell types, their sensitivity to pharmacological and genetic manipulations, and their resulting impact on behavior.

* Second, we use higher-level mathematical models to analyze the functional properties of the neurocognitive systems in terms of reinforcement learning (RL), hierarchical Bayesian inference, and sequential sampling models of decision-making, affording principled computational interpretations and allowing for tractable quantitative fits to brain/behavior relationships.  Various efforts in our lab involve linking the mechanistic models to interpret their function within normative frameworks and statistically estimable and parsimonious low dimensional models.

* Third, we apply “computational multidimensional functional profiling”. This approach adopts ML methods that combine features extracted from theory-driven models (e.g., latent parameters) with data-driven approaches, to discover clusters of features maximally predictive of relevant phenotypes and brain states.

For empirical support, we characterize neural mechanisms of human and rodent behavior. We design computerized tasks sensitive to neural computations that probe reinforcement learning, cognitive control, and decision-making under uncertainty. We provide

quantitative estimates of individual performance parameters using mathematical models that assess the degree to which subjects rely on specific computations when learning and making decisions, and how they vary with neural activity and genetics. Our laboratory has primarily focused on human experiments (pharmacology, intracranial neural recordings from patients, EEG, fMRI, genetics), but we also participate via collaboration in rodent research (genetic knockouts, optogenetics, calcium imaging, electrophysiology and pharmacology), affording detailed probing of model mechanisms.

A neural network of the basal ganglia in reinforcement learning approximates a bayesian learner. This model simulates pause patterns in striatal cholinergic neurons, their interactions with dopamine and medium spiny neurons, and their impact on learning rate in stochastic and dynamic environments. This model suggests that the basal ganglia can alter its own learning rate as a function of its uncertainty about task contingencies (variable pauses), while remaining sensitive to changes in those contingencies as in reversal tasks, approximating a Bayesian learner. From Franklin & Frank (2015), eLife