The geometry of neural representations reflecting abstraction in humans

ABSTRACT The process of abstraction involves identifying the features shared across past experiences so as to represent a complex environment using only a small number of variables. Abstraction obviates the need to represent all combinations of values for all features and enables generalization to novel environments. Such generalization is

fundamental to rapid, flexible adjustments in behavioral, cognitive, and emotional responses. However, it re- mains unknown how the human brain learns to represent past experiences to reflect their shared features and enable generalization, nor how this process is modulated by different timescales of learning, memory consolida-

tion, multiple levels of abstraction, and motivational states. To answer these questions, we adapt for human fMRI a theoretical framework and analytic methodology from recent work in non-human primates. Healthy human subjects learn a complex reversal-learning task with multiple stimuli linked by a hidden structure that can be

represented by a small number of variables. In pilot data, subjects learn this structure, which they demonstrate via inference: a change in one stimulus is sufficient to infer the new values for the remaining stimuli. We analyze multivoxel fMRI activity to probe for the relationships, or geometry, between neural representations, which we

test for an ‘abstract format’, i.e., a format that enables generalization, as well as quantify its dimensionality, or capacity to represent a large number of (non-abstract) variables. In Aim 1, we probe the evolution of neural representations during learning at multiple timescales, from hours to a week, to provide mechanistic insight into

the formation and consolidation of abstract representations. We predict that (1) the abstract format will emerge first for ‘explicit’ variables (e.g., response and outcome) in regions associated with sensorimotor processing. (2) After multi-day training and consolidation, we predict that ‘hidden’ variables defined by the task’s temporal

statistics will be represented in an abstract format, first by regions that encode relational knowledge (e.g., medial temporal lobe), which then relay this information to prefrontal regions that encode abstract rules and task states. Aim 2 compares different levels of abstraction, from identifying shared features across specific instances to a

system of general states that can be transferred to novel problems. We compare the neural geometry and brain regions (e.g., hippocampus vs. entorhinal cortex) that support these distinct levels. Aim 3 investigates the role of appetitive vs. aversive outcomes, which profoundly influence decision-making and learning, but their distinct

roles in abstraction are unknown. This gap is striking given that many psychiatric disorders involve impaired abstraction and generalization tied to aversive experiences. To address this gap, subjects perform alternating versions of the task under gain or loss domains. We will test how motivational valence impacts abstract learning

and neural geometry. In all Aims, we relate individual differences in behavior and affective processing to differences in neural geometry. Going forward, the framework provides a foundation for linking neural geometry to cognitive and emotional function with broad applicability to cognitive neuroscience and psychopathology.