Contrasting Success in Neurotechnology: Sensory Substitution, Brain–Computer Interfaces, and the Limits of Dimensional Reduction

Neurotechnologies interfacing with sensory and motor systems have achieved markedly different outcomes despite pursuing similar restorative goals. Sensory substitution devices—which convert visual information to tactile or auditory signals with three orders of magnitude less spatial resolution—enable users to navigate, recognize objects, and perform complex visuospatial tasks. In contrast, brain–computer interfaces for motor control remain limited despite high-density neural recording and sophisticated machine learning decoding. This essay examines this divergence by analyzing how information reduction affects perceptual versus neural systems differently. Recent large-scale neural recordings (up to one million neurons) reveal power-law scaling of dimensionality with population size, where fine-scale but functionally relevant dimensions emerge only at high sampling density. We argue this finding has profound implications for neurotechnology design: dimensional reduction techniques that work effectively for artificial systems may fundamentally misrepresent biological neural organization. Meanwhile, sensory substitution succeeds precisely because perceptual systems can extract behaviorally relevant structure—what ecological psychology describes as affordances—from dramatically coarse-grained input across different forms of stimulation. Drawing on ecological psychology and recent dimensionality studies in neuroscience, we propose that effective neurotechnology must respect this asymmetry. Sensory interfaces should preserve ecological structure rather than maximize information content, while neural decoding must accommodate scaling complexity rather than impose dimensional reduction. We discuss specific implications for cortical visual prostheses, cochlear implants, and bidirectional brain–computer interfaces, arguing that theoretical frameworks treating perception and neural activity as unified computational processes may hinder rather than advance neurotechnology development.