The neural bases of adaptive behavior in social environments are far from being understood. We propose to use both computational and neuroscientific methodologies to provide new and more accurate models of learning in interactive settings. The long-term objective is to develop a neural theory of learning: a mathematical framework that describes the computations mediating social learning in terms of neural signals, structures and plasticity. We plan to develop a model of adaptive learning based on three basic principles: (1) the observation of the outcome of un-chosen options improves the decisions taken in the learning process, (2) learning can be transferred from one domain to another, and (3) learning can be transferred from one agent to another (i.e. social learning). In all three cases, humans appear able to construct and transfer knowledge from sources other than their own direct experience, an underappreciated though we believe critical aspect of learning. Our approach will combine neural and behavioral data with computational models of learning. The hypotheses will be formalized into machine learning algorithms and neural networks of “regret” learning, to quantify the evolution of the learning computations on a trial-by-trial basis from the sequence of stimuli, choices and outcomes. The existence and accuracy of the predicted computations will be then tested on neural signals recorded with functional magnetic resonance imaging (fMRI). The potential findings of this project could lead us to suggest general principles of social learning, and we will be able to measure and model neural activation to show those general principles in action. In addition, our results could have important implications into policy-making - by revealing what type of information agents are naturally inclined to better learn from - and clinical practice - by outlining potential diagnostic procedures and behavioral therapies for disorders affecting social behavior.