Neural representation of behavioral outcomes in the orbitofrontal cortex

The orbitofrontal cortex (OFC) is important in processing rewards and other behavioral outcomes. Here, we review from a computational perspective recent progress in understanding this complex function. OFC neurons appear to represent abstract outcome values, which may facilitate the comparison of options, as well as concrete outcome attributes, such as flavor or location, which may enable predictive cues to access current outcome values in the face of dynamic modulation by internal state, context and learning. OFC can use reinforcement learning to generate outcome predictions; it can also generate outcome predictions using other mechanisms, including the evaluation of decision confidence or uncertainty. OFC neurons encode not only the mean expected outcome but also the variance, consistent with the idea that OFC uses a probabilistic population code to represent outcomes. We suggest that further attention to the nature of its representations and algorithms will be critical to further elucidating OFC function.

Introduction

The orbitofrontal cortex (OFC) was initially characterized as an area whose destruction profoundly impacted human personality, but, paradoxically, left no obvious deficits in standard cognitive tests (reviewed in [1]). Yet, through intensifying scrutiny over the last decade the function of the OFC has arisen from obscurity to take a central place in our understanding of learning and decision-making [2, 3]. Today, through a remarkable convergence of studies conducted in species ranging from rats to humans, OFC is widely conceived as a place where the ‘value’ of things is represented in the brain.

While the concept of ‘value’ may strike a hard-nosed neuroscientist as hopelessly fuzzy, this concept plays a central role in most behavioral theories of decision-making. In neuroeconomic theory, assignment of economic value allows qualitatively different goods to be compared in a single ‘universal currency’ [4]. In animal learning theory, the similar concept of ‘incentive value’ measures the ability of outcomes to motivate behavior [5, 6]. In machine learning theory, ‘state values’ and ‘action values’ are the principal targets of learning and action selection; by maximizing these values, agents learn optimal behavior [7]. By offering formal (i.e. quantitative) definitions of value and related concepts, these theoretical frameworks can help one to test and eventually to understand more precisely what the OFC does. That is because formal definitions can yield concrete predictions that are testable using traditional neurophysiological and behavioral measurements without resorting to semantic arguments about abstract terms [8].

While theoretical perspectives are helpful, they also bring on more work. In the light of theory, questions about OFC function become not only more clear but also more detailed and nuanced, opening up and demanding further experimental tests. Moreover, different theoretical frameworks present partially overlapping, sometimes incongruent, views that must eventually be reconciled. Finally, applying theories of behavior to the brain requires one to bridge the gap between the functional level that forms the basis for the theory and the level of neurophysiology. As famously framed by David Marr [9], two key pieces are needed to bridge between behavioral (computational) and neural (implementational) levels: first, understanding the nature of the neural code or representation; second, understanding the processes or algorithms used to create and utilize these representations.

This review will examine recent progress in OFC function in light of economic, psychological and computational theories of value. While we wholeheartedly acknowledge the convergence of many threads evidence, our main goal is to emphasize the ragged edges and emerging complexities. These become apparent especially when asking what exactly a neural representation within OFC might look like, and therefore our primary focus will be on recordings from individual OFC neurons in monkeys and rats, with secondary attention to lesions and neuroimaging studies. We will also review what we know about the origin of these representations, and touch upon the issue of how they are used.