Abstract: The recent decade has witnessed an explosive development of the family of predictive mind theories stemmed from Karl Friston's Free Energy Principle (FEP), a family in which there exist two main streams: the Predictive Coding theory (PC) defended by Jacob Hohwy and the Predictive Processing theory (PP) advocated by Andy Clark. The former holds a standpoint of the brain which Hohwy deemed as the only source and center of cognition, and the latter emphasize the viability of the whole organism defined by Clark as the only goal and criteria of mind. We attempt to make a deep dive into the theoretical substrate of the predictive mind and build a bridge between PP and enactivism with the bricks of (basic) dynamical system theories.
This work has revealed that PP in essential is a representative theory of systemic dynamics, and has not indeed appointed the brain, the brain-body, or the brain-body-environment as the only bearer of its basic architecture whose only mission is to minimize prediction error. Prediction Error Minimization (PEM) via the coupling of generative model and generative process by implementing active inference can happen at multiple spatial and temporal scales, the expression of which include the morphology, action tendency and gross neural architecture of adaptive organisms. This nature of PP is determined by FEP, which, although in itself stems from theoretical neural science researches by Friston and his colleagues, attempts to play a role of Hamilton's principle of minimal action in cognitive science by presenting a basic dynamical rule of all life and mind aimed at fighting against the random fluctuation and second law of Thermodynamics to preserve their own integration and distinction. With the conceptual framework of FEP and dynamical system theories, predictive processing and PEM can happen at any scale where there exist two circular causal loops, and a dynamical coupling between an agent and its environment can viably endure. It is this coupling that enact the agency and adaptive activities, separating the agent and its environment in an inferential sense but integrating them existentially. Thus, it is not the viable organism that seeks to minimize prediction error by engaging the environment but the homeostatic system as a whole that entail the endurance of the viable relationship between the constituents of the system.
This systemic standpoint has a deep resonance with enactivism, which viewed cognition as an emergent phenomenon from activities of life in niche. The three core concepts of PP, respectively PEM, generative model and active inference has an elegantly clear correspondence with the three main groups of enactivism, that is, autopoiesis enactivism, sensorimotor enactivism and radical enactivism. For autopoiesis enactivism, sense-making can be translated into the attunement of the dynamical system with the quantitative measurement of PEM, and it is the systemic dynamics rather than organism dynamics that determine the prescriptive rules which guarantee the strong continuity of life and mind. For sensorimotor enactivism, it is the SMCs that characterize the phenomenology of perceptual experience and the SMCs knowledge in essential is the grasp of the if-then relationship between the exploratory activities and the resulting perception, and ultimately comes from the coupling of generative model and generative process. For radical enactivism, content of higher-level cognitive phenomenon comes from engaging the designer environment, namely the cultural framework embodied as the abundant material symbolic environment self-generated and self-reproduced. All in all, enactive PP holds a standpoint of the dynamical systems, emphasizes the whole system over organism or agent, model over representation, designer environment over objective environment, and has a prospect of clarifying the source of sense of cognition, the phenomenology of perceptual experience and the dynamics of the mind.

Key words: predictive processing, enactivism, dynamics, system

摘要： 预测加工理论主张大脑是一台多层概率预测引擎,致力于最小化预测误差,主要策略包括：（1）修正先验以改变预测;（2）采取行动以改变输入。尽管预测加工理论依然预设了内部表征,但其作为一种动力学理论同时具有鲜明的生成主义色彩。从底层系统动力学出发,对比预测加工理论与认知科学的生成主义取向,可知“生成主义预测加工理论”有望推动认知科学的“动力学化”,揭示认知的意义来源、知觉的现象学与心智的动力学。

关键词: 预测加工, 生成主义, 动力学, 系统