On Organizational Principles of Neural Systems

Xin Li

2024-02-22

Abstract: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.

Neurons and Cognition

What problem does this paper attempt to address?

This paper attempts to explore the organizational principles of the nervous system from three levels (device/implementation layer, circuit/algorithm layer, and system/computation layer) in order to understand natural intelligence. Specifically:

1. **Device/Implementation Layer**: The paper studied the reproducibility and scalability of macroscopic states in polysynchronous neural groups (PNGs), which form the physical basis of associative memory. Through the spike - time - dependent plasticity (STDP) - based time - locking mechanism, PNGs can reduce Kolmogorov - Sinai entropy, thus promoting the reproducibility of macroscopic states. In addition, the time - locking mechanism of multiple PNGs can be scaled in a way similar to that of a single neuron.

2. **Circuit/Algorithm Layer**: The paper proposed that predictive coding is a general cortical processing algorithm for modeling basic sensorimotor interactions. The predictive processing of canonical micro - circuits can be regarded as the generalization of associative memory in cortical columns. At the cortical column level, sensorimotor interactions can be abstractly represented by a simultaneous localization and mapping (SLAM) model based on Kalman filtering.

3. **System/Computation Layer**: The paper emphasizes that behavioral flexibility is the basis of intelligent systems. By connecting the sensory and motor systems, the neocortex generates adaptive behaviors by minimizing prediction errors. Using coordinate transformations (e.g., from egocentric to allocentric), action - oriented representations can navigate and recall associative memories in the latent space, thereby solving various high - level cognitive problems. For example, visual and language models can be understood as spatial navigation in the latent space defined by coordinate transformations.

The main contribution of the paper lies in proposing a unified framework that explains how agents evolve through adapting to complex behaviors in the physical environment and social groups. This framework combines dynamic systems theory and predictive coding, aiming to implement the general processing algorithm proposed by Vernon Mountcastle. In this way, the paper hopes to provide a new perspective for the cross - research between neuroscience and learning systems and bridge the gap between natural intelligence and artificial intelligence.