Bio-Inspired Neural Network Architecture of Embodied Intelligence

Advancements in artificial intelligence (AI) and machine learning over the past years have been driven by the evolution of large language models (LLM). However, even such models as ChatGPT-4o have considerable drawbacks which significantly limit the areas of their applicability. One of the reasons for LLMs limitations is their lack of knowledge acquired through implicit learning, which is key for any biological organism. Embodied intelligence is the main form of implicit knowledge acquisition, which is required for solving interactive physical problems of the real world. Bio-inspired studies of the nervous systems that organisms have allow us to consider the cerebellum, which coordinates movements and maintains balance, as a major way to uncover methods for embodied physical intelligence development. Its simple repetitive structure and ability to control complex movements hold promise of creating an analog to adaptive neural networks. This paper explores the biologically-inspired architecture of the cerebellum as an analog computational network capable of modeling complex physical systems in the real world. The model of an octopus tentacle taken as a simple example illustrates how the embodied AI, which integrates sensory perception and physical interaction, can be implemented. The model demonstrates the potential for creating adaptive physical systems that learn and interact with the environment.