Editorial: Self-Organization in the Nervous System

摘要

“Self-organization is the spontaneous—often seemingly purposeful—formation of spatial, temporal, spatiotemporal structures, or functions in systems composed of few or many components. In physics, chemistry and biology self-organization occurs in open systems driven away from thermal equilibrium” (Haken, Scholarpedia). The contributions in this special issue aim to elucidate the role of self-organization in shaping the cognitive processes in the course of development and throughout evolution, or “from paramecia to Einstein” ( Torday and Miller ). The central question is: what self-organizing mechanisms in the human nervous system are common to all forms of life, and what mechanisms (if any) are unique to the human species? Over the last several decades, the problem of self-organization has been at the forefront of research in biological and machine intelligence (Kohonen, 1989 ; Kauffman, 1993 ; Pribram, 1994 , 1996 , 1998 ; Kelso, 1997 ; Camazine et al., 2003 ; Zanette et al., 2004 ; Haken, 2010 , 2012 , and others). The articles collected in this issue present recent findings (and ideas) from diverse perspectives and address different facets of the problem. Two features of this collection might be of particular interest to the reader: (i) the scope of discussion is broad, stretching from general thermodynamic and information-theoretic principles to the expression of these principles in human cognition, consciousness and understanding and (ii) many of the ideas speak to a unifying perspective outlined below. In what follows, we will preview the collection of papers in this special issue and frame them in terms of a unified approach to self organization—leaving the reader to judge the degree to which subsequent articles are consistent with or contradict this framework. Living organisms must regulate flows of energy and matter through their boundary surfaces to underwrite their survival. Cognitive development is the product of progressive fine-tuning (optimization) of regulatory mechanisms, under the dual criteria of minimizing surprise (Friston, 2010 ; Sengupta et al., 2013 , 2016 ; Sengupta and Friston, 2017 ) and maximizing thermodynamic efficiency (Yufik, 2002 , 2013 ). The former implies reducing the likelihood of encountering conditions impervious to regulation (e.g., inability to block inflows of destructive substances); the latter implies maintaining net energy intakes above some survival thresholds. Energy is expended in regulatory processes formed in the course of self-organization and predicated on lowering thermodynamic entropy “on the inside” and transporting excessive entropy (heat) “to the outside.” Efficient regulation requires mechanisms that necessarily incorporate models of the system and its relation to environment (Conant and Ashby, 1970 ). Primitive animals possess small repertoires of genetically fixed, rigid models, while—in more advanced animals—the repertoires are larger and their models become more flexible; i.e., amenable to experience-driven modifications. Both the evolutionary and experience-driven modifications are forms of statistical learning: models are sculpted by external feedback conveying statistical properties of the environment. Human learning mechanisms, although built on the foundation of statistical learning, depart radically from conventional (e.g., machine) learning: the implicit models become amenable to self-directed composition and modification based on interoceptive, as opposed (or in addition) to exteroceptive, feedback (Yufik, 1998 ). Interoceptive feedback underlies the feeling of grasp, or understanding that accompanies the organization of disparate “representations” into cohesive structures amenable to further operations (mental modeling). The work of mental modeling requires energy; consciousness is co-extensive with deliberate (attentive, focused) application of energy (“cognitive effort”) in carrying out that work. Learning with understanding departs from statistical (machine) learning in three ways: (i) mental models anticipate experiences, as opposed to be shaped by them (e.g., the theory of relativity originated in gedanken experiments); (ii) feedback conveys properties of implicit models (coherence, simplicity, validation opportunities the models afford, etc.) and (iii) manipulating (executing or inverting) models enables efficient exchange with the environment, under conditions with no precedents (and thus no learnable statistical representation) (Yufik, 2013 ). Regulation of this sort—based on statistical learning—faces a challenging complexity. As the number of regulated variables grows; energy demands can quickly become unsustainable. Using self-organization to implement the process of “understanding” (i.e., composing more general models) has the triple benefit of minimizing surprise, while averting complexity and advancing thermodynamic efficiency of regulatory processes into the vicinity of theoretical limits. Annila argues that the most funda