Neurophysiology underlying neural representations of value

Understanding the neural mechanisms underlying decision-making is important because patients with many psychiatric disorders make mal-adaptive decisions, impacting executive functioning, including emotion and mood regulation. Historically, the mechanisms underlying decision-making have been most studied using behavioral paradigms in which subjects repeatedly make decisions about well-controlled stimuli or options, invoking perceptual, valuation, memory, or other processes. These studies have provided significant insight, but many if not most decisions in the real world occur in very different circumstances than those realized in the laboratory. Two paradigmatic types of such decisions are those made in novel situations never encountered before, and those that require subjects to generate new responses to familiar stimuli, i.e. to flexibly adjust behavior in a new way. In this grant, we test the hypothesis that mechanisms underlying these 2 forms of decision-making can be revealed by examining the 'geometry' of neural representations and relating them to behavior in tasks invoking the 2 types of decisions. The geometry of a representation is defined by the set of all distances between points in the activity space that represent responses of multiple neurons in different conditions. Measures of a representational geometry include assessment of its dimensionality. Decisions in novel situations require generalizing from past experiences to a new one, an ability relying on abstraction. Abstraction constructs variables describing features shared by instances within and across situations, capturing regularities and structure in the world. Neural representations of abstracted variables are similar to the widely studied disentangled representations in machine learning and have lower dimensionality. On the other hand, high dimensional neural representations support the ability to generate many different responses without changing the underlying representation. Recent data indicate that neural ensembles in the hippocampus (HPC) and prefrontal cortex (PFC) achieve geometries with a sufficiently low dimensional ‘scaffold’ to support generalization in new situations, but the scaffold is embedded in a higher dimensional representation of task variables. This geometry has specific computational capabilities, but do they actually relate to decision-making? Here we combine high-channel count electrophysiology, neural network modeling, and carefully designed tasks to provide evidence for the first time that these 2 aspects of the geometry actually are used to support the 2 distinct types of decision-making. We examine the geometry of representations in HPC and PFC in relation to decisions of both types. We ask if the ability to make decisions relying on abstraction correlates with how a key task-relevant variable is represented in a low-dimensional scaffold (Aim 1). Then we test if the representation encodes many other variables with higher dimensionality, and if this predicts the ability to learn to generate new responses (Aim 2). Finally, we train neural network models on the same tasks to elucidate neural coding principles in HPC and PFC, and interactions between areas mediating these 2 forms of decision-making (Aim 3).