사회과학을 위한 액티브 인퍼런스

There is no clear-cut distinction between proprioception and exteroception — or rather, any such distinction relies on a specific model of what constitutes “me.” An embryo, for example, is treated by its mother as part of her body; after birth it is no longer directly a component of the mother's metabolism, though it remains embedded within the mother's self-model for a time, and only gradually becomes fully disembedded from it psychologically, even though birth is a clear metabolic cutting point. So whether the mother-child relation is “interoceptive” is genuinely unclear — what is clear is that a rigid conception of either interoception or exteroception will lead to prediction error. The most directly relevant paper is Ciaunica, Constant, Preissl, and Fotopoulou (2021), “The First Prior: From Co-Embodiment to Co-Homeostasis in Early Life,” in Consciousness and Cognition; the notions of “co-embodiment” and “co-homeostasis” it develops apply more widely to social groups, with different forms of coordination than in mother-child dyads. More broadly, human cognition is a permanent dance between internalizing and externalizing systems, part of one's extended phenotype or part of the environment being navigated; there is no “true” scale of individuation, only a scale of agency relative to specific perception-action cycles.

There is no straightforward answer to this question. The free-energy principle entails a total symmetry between an agent and its environment — the environment (taken as a whole) is “self-evidencing” with the agent as its environment. A good theory of cognition would integrate what enactivists call “interactional asymmetry”: the fact that the agent has the exclusive ability to define the terms of the interaction, which resonates with the FEP requirement that “external states” are states the agent can hold beliefs over — states that are cognitively meaningful to it. Artefacts likely constrain an agent's self-model in a way that constitutes normativity (see Guénin-Carlut and Albarracin, 2023, “On Embedded Normativity — An Active Inference Account of Agency beyond Flesh,” OSF Preprints), opening the door to a full model of embedded, distributed agency — without artefacts being agentive in and of themselves, but rather serving as means through which bona fide agents coordinate. The formal grounding for this is still developing; see also Guénin-Carlut (2022), “Physics of Creation — Symmetry Breaking, (En)Active Inference, and Unfolding Statespaces,” OSF Preprints.

Precision, within Active Inference, is understood as the confidence of a given prediction — precisely, the inverse of its variance. A good reference is Feldman and Friston (2010), “Attention, Uncertainty, and Free-Energy,” in Frontiers in Human Neuroscience.

A rapid search surfaced papers where Solms argues that the predictive mind largely explains away consciousness (see Solms 2019, “The Hard Problem of Consciousness and the Free Energy Principle,” Frontiers in Psychology, and Solms 2020, “New Project for a Scientific Psychology: General Scheme,” Neuropsychoanalysis), which would need to be weighed more specifically against the particular arguments the questioner has in mind as compelling evidence against prediction underlying emotion.

The course did not propose pre-defined topics, but the organizers were open to discussing individual topic ideas directly with interested participants through the course's contact channels.

The general approach is to model collective dynamics by modeling individual systems as prediction-error minimizers — a more informative, though formally equivalent, approach to a classical dynamical-systems treatment. See Heins, Millidge, da Costa, Mann, Friston, and Couzin (2023), “Collective Behavior from Surprise Minimization,” arXiv.

This question was submitted for discussion; a full course response was not recorded in the distributed material.

Enactivists have addressed this in some form, moving from biological individuation toward a theory of ethics — see Varela (1999), Ethical Know-How: Action, Wisdom, and Cognition (Stanford University Press), and, on effortless virtue, Slingerland (2014), Trying Not to Try (Canongate Books). A recurring tension in the discussion: normative claims carry goals (directive or aspirational), so “what is” cannot straightforwardly ground “what ought to be” — nature is not itself ethically motivated, even though biological individuation is grounded in the autopoietic basis of the agent. Much of the existing work here is descriptive (how ethics and norms actually distribute across populations) rather than normative, and hard positive (prescriptive) results are correspondingly hard to derive; one candidate intuition raised was that distributed empowerment tracks “goodness,” in the sense that a system occupying its most-expected place is doing well by its own lights, echoing the Panglossian “best of all possible worlds” framing against a Bayesian “most likely world.”

Any model requires choosing what to include and exclude — initial and boundary conditions, and how the relevant state space is specified, given that novel state spaces can themselves arise over time (see Guénin-Carlut, 2022, “Cognitive Agency in Sociocultural Evolution,” OSF Preprints). These general modeling constraints apply to any framework. Active Inference's comparative advantage lies in how it specifies constraints and explanations: it captures a duality between informational constraints and the beliefs, intentions, and desires an organism holds, in a way that purely non-representational dynamics (e.g. Hebbian learning, where no “beliefs” or “norms” emerge as non-trivial causal entities) and purely representational frameworks (e.g. classical utility theory, which must pre-specify intentional states) cannot, allowing normativity and intentionality to be specified indirectly rather than assumed. This connects to a broader question of which formal “languages of nature” (message passing, quantum information, category theory) are complementary and non-reducible ways of describing the same phenomena — see Baez and Stay (2011), “Physics, Topology, Logic and Computation: A Rosetta Stone,” in New Structures for Physics; Fuchs and Schack (2011), “A Quantum-Bayesian Route to Quantum-State Space,” Foundations of Physics; and Felin, Kauffman, Koppl, and Longo (2014), “Economic Opportunity and Evolution: Beyond Landscapes and Bounded Rationality,” Strategic Entrepreneurship Journal.

Session slides, where available, were attached directly to each session in the course's syllabus; the Active Inference Journal also curates course-related transcripts and reading material.