On Organizational Principles of Neural Systems

How do we understand natural intelligence? Inspired by classical embodied cognition and the emerging multimodal interaction, we study the organizational principles of neural systems at three levels (device/implementation, circuit/algorithm, and system/computational) in this survey paper. Our main contributions consist of 1) Device/Implementation level: reproducibility of macroscopic states in polychronization neural group (PNG) serves as the physical basis of associative memory; 2) Circuit/Algorithm level: canonical microcircuits implement a universal predictive coding algorithm underlying all high-level cognitive functions; 3) System/Computational level: modeling sensorimotor interaction for embodied cognition plays a fundamental role in understanding natural intelligence. At each level, we use mathematical models as our abstractions and study their organizational principles (e.g., entropy reduction, predictive coding, and coordinate transformation). The unifying theme is that natural intelligence evolves by recycling a simple navigation trick in different coordinates (reference frames) hundreds of thousands of times. By representing space as a latent sequence, the neocortex solves the embodied cognition problem by building parallel cognitive maps at a large scale. We hope this survey article can inspire new research at the intersection of neuroscience and learning systems, helping bridge the gap between natural and artificial intelligence.