Why Meaning Survives Noise: The Spatiotemporal Abstraction Theory

Abstract

The brain excels at extracting meaning from noisy and degraded input, yet the computational principles that underlie this robustness remain unclear. We propose a theory of spatiotemporal abstraction (STA), in which neural networks integrate inputs across space and time to produce multi-scale, concept-level representations that remain stable despite loss of detail. We demonstrate this principle using spectrograms of spoken sentences and their degraded analogs from cochlear implants, showing that as integration kernels widen, distorted input converges toward the original representation. This mechanism may explain how cochlear implant users comprehend speech despite severely scrambled afferent signals. STA provides a unified framework for understanding abstraction as an emergent property of cortical architecture, with implications for memory, neuroprosthesis design, and robust artificial systems.