Direct codes in a distributed age.
TL;DR
It is suggested that the central challenge is not to choose between direct and distributed codes, but to determine when geometric descriptions illuminate mechanism and when they merely rename it.
Abstract
Large-scale recordings and state-space analyses have transformed systems neuroscience by revealing low-dimensional structure in neural population activity. Yet these advances also invite a risk: they can obscure the neurobiological organization that gives rise to computation. I revisit Horace Barlow's distinction between direct and distributed coding in light of modern work on perceptual decision-making. The starting point is an older idea: cortical signals are noisy, and this noise implies pooling across neurons with similar response properties. In this view, variables are represented directly in the firing rates of identifiable neurons or neuronal pools. Modern population analyses often recover the same variables as projections along coding directions in a high-dimensional state space, creating the appearance of distributed representation. I argue that these two descriptions are not mutually exclusive, but they are not equivalent. In some cases, a distributed description re-describes a computation that is implemented in a more direct and mechanistically interpretable way by specific neural subpopulations. Using work on decision-related activity in the lateral intraparietal cortex and superior colliculus as a case study, I suggest that the central challenge is not to choose between direct and distributed codes, but to determine when geometric descriptions illuminate mechanism and when they merely rename it.