On Computational Modeling of Sleep-Wake Cycle (2404.05484v2)

Abstract: Why do mammals need to sleep? Neuroscience treats sleep and wake as default and perturbation modes of the brain. It is hypothesized that the brain self-organizes neural activities without environmental inputs. This paper presents a new computational model of the sleep-wake cycle (SWC) for learning and memory. During the sleep mode, the memory consolidation by the thalamocortical system is abstracted by a disentangling operator that maps context-dependent representations (CDR) to context-independent representations (CIR) for generalization. Such a disentangling operator can be mathematically formalized by an integral transform that integrates the context variable from CDR. During the wake mode, the memory formation by the hippocampal-neocortical system is abstracted by an entangling operator from CIR to CDR where the context is introduced by physical motion. When designed as inductive bias, entangled CDR linearizes the problem of unsupervised learning for sensory memory by direct-fit. The concatenation of disentangling and entangling operators forms a disentangling-entangling cycle (DEC) as the building block for sensorimotor learning. We also discuss the relationship of DEC and SWC to the perception-action cycle (PAC) for internal model learning and perceptual control theory for the ecological origin of natural languages.