关键词: 预测加工, 层级预测编码, 自由能原理, 主动推断

Abstract: Nowadays, quarrels and fights between two generations of cognitive science almost result in an inevitable impasse. As an instance, the deep relations of theories between J. J. Gibson and David Marr has been carved off, and of which two theories each other has been pushed toward the opposite sides; which could be viewed as a portrayal of these overt states of affairs. The central tenet of the First-Generation Cognitive Science summarizes as: the Representational-Computational Approach; and the central tenet of the Second-Generation Cognitive Science: the Embodied Action Approach. What is the nature of “mind” at all? These two approaches conflict each other; and the either/or choice which seems the only game in the town. In recent years, researchers attempt to integrate both strains of sources, hoping to provide a unified framework, thus of whichc ould be viewed as might prospectively break the impasse of two generations of cognitive science in the near future. It calls “Predictive Processing”. This approach apparently originates in hierarchal predictive coding theories of vision & auditory studies (since 1980s), and thereby has been modified substantively by the neuroscientist Karl Friston, et al., into a fashion of action-oriented. Predictive Processing combines with two indispensable components: Hierarchical Predictive Processing and Active Predictive Processing. In rough, the Hierarchical Predictive Processing mainly inherits from hierarchical predictive coding theories of the First-Generation Cognitive Science(e.g., Marr’s multi-level computational theory of vision, hierarchical neural networks, etc.); and the Active Predictive Processing somehow advances various versions of theories of the Second-Generation Cognitive Science relates with action(e.g., Gibson’s affordances theory, enactive cognition, etc.). Both components could be viewed as dual aspect of an integrated theory. In Predictive Processing, the Free Energy Principle views adaptivity as the most fundamental principle of the theory construction of neural & cognitive systems, and generally self-organizing systems. The overall goal of Hierarchical Predictive Processing is to make the cognitive agency adaptable and sustainable as much as possible. Next, the optimization problem of adaptation turns into problems of surprises minimization or prediction errors minimization by means of Bayesian inference & Bayesian statistics. The basic working mechanism of Hierarchical Predictive Processing: ascending prediction-descending prediction errors tradeoffs. The so-called “Active Inference” which is the complementary components of Free Energy Principle; And which be seen by Paul Cisek et al., is confirmed with multi-affordances competition mechanisms. The Active Predictive Processing shows perception-action tradeoffs, or “epistemic-pragmatic tradeoffs”. Both perception (epistemic value) and action (pragmatic value) fundamentally serves for adaptivity (existential value). Perception-action cycle of Predictive Processing which could be regarded as a second-order cycle based on first-order embodied reflex arc; which is also internalized as a closed loop in the head. Thus on the conception of cognition, contrasting with past two generations of cognitive science, Predictive Processing may has made revolutionary breakthrough at all. Cognition as the systemic function of general adaptive system complies with the most fundamental adaptivity principle. The conventional object-invariants analysis beneath paradigms of two generations of cognitive science, thus ought to give way to mind-world adaptivity analysis underlying the Predictive Processing paradigm. Philosophically speaking, Predictive Processing seems accords with a process theory. In the end, on the basis of the currents and trends of Predictive Processing, we attempt to propose an integrated view. In our view, the pressing contemporary task of Predictive Processing is trying to integrate two generations of cognitive science—Representational-Computational Approach and Embodied Action Approach.

Key words: predictive processing, hierarchical predictive coding, free energy principle, active inference