Effective learning is accompanied by increasingly efficient dimensionality of whole-brain responses

Theories and tools to measure the efficiency of neural codes have been important in understanding neural responses to external stimuli. However, similar principles to describe the efficiency of brain responses to tasks demanding higher-order cognitive processes remain underdeveloped. A key domain to study such efficiency is learning, where patterns of activity across the entire brain provide insight into the principles governing the acquisition of knowledge about objects or concepts. We propose a mathematical framework for studying the efficiency of spatially embedded whole-brain states reflected in functional MRI, and demonstrate its utility in describing how human subjects learn the values of novel objects. We find that compared to slow learners, quick learners use a smaller ambient embedding to simultaneously encode higher dimensional patterns of neural responses to stimuli, providing a notion of effective coding most associated with rapid learning. Furthermore, we identify which regions form the strongest neurophysiological drivers of these differences in learning rate, and complement our region-based approach with a voxel-level approach to uncover structure in finer-scale responses. Finally, for a full investigation of complementary geometric approaches, we verify that quick learners develop more assortative responses to stimuli: that is, whole-brain responses that are more easily distinguishable from one another. Our work offers a suite of geometric measures to represent a notion of efficient coding for higher-order cognitive processes, and provide insight into the dimensionality of neural responses characteristic of the successful optimization of reward, that are applicable to the study of cognitive performance more broadly.