Integrated World Modeling Theory (IWMT) Implemented: Towards Reverse Engineering Consciousness with the Free Energy Principle and Active Inference

Abstract

Integrated World Modeling Theory (IWMT) is a synthetic model that attempts to unify theories of consciousness within the Free Energy Principle and Active Inference framework, with particular emphasis on Integrated Information Theory (IIT) and Global Neuronal Workspace Theory (GNWT). IWMT further suggests predictive processing in sensory hierarchies may be well-modeled as (folded, sparse, partially disentangled) variational autoencoders, with beliefs discretely-updated via the formation of synchronous complexes—as self-organizing harmonic modes (SOHMs)—potentially entailing maximal a posteriori (MAP) estimation via turbo coding. In this account, alpha-synchronized SOHMs across posterior cortices may constitute the kinds of maximal complexes described by IIT, as well as samples (or MAP estimates) from multimodal shared latent space, organized according to egocentric reference frames, entailing phenomenal consciousness as mid-level perceptual inference. When these posterior SOHMs couple with frontal complexes, this may enable various forms of conscious access as a kind of mental act(ive inference), affording higher order cognition/control, including the kinds of attentional/intentional processing and reportability described by GNWT. Across this autoencoding heterarchy, intermediate-level beliefs may be organized into spatiotemporal trajectories by the entorhinal/hippocampal system, so affording episodic memory, counterfactual imaginings, and planning.

Back to Square One: The Bodily Roots of Conscious Experiences in Early Life

“…IWMT has previously focused on GNWT and IIT, but its ultimate goal is creating a framework for synergistically combining all other theories of consciousness as sources of complementary perspectives and converging lines of evidence, as well as helping to adjudicate between conflicting truth claims. IWMT grounds itself in the Free Energy Principle and Active Inference (FEP-AI) framework as a general systems theory and normative model of adaptive/intelligent functioning (Friston, 2010;Friston et al, 2017), as well as the associated computational process theory of (hierarchical) PP (Bastos et al, 2012;Safron, 2020b).…”

Integrated World Modeling Theory (IWMT) Expanded: Implications for Theories of Consciousness and Artificial Intelligence

“…Further conceptual and functional integration between AIXI and FEP-AI could potentially be achievable through the study of generative modeling based on probabilistic programs (Lake et al, 2015;Lázaro-Gredilla et al, 2019;Ullman & Tenenbaum, 2020). With relevance to IWMT, consciousness may be understood as kind of lingua franca (or semi-interpretable shared latent space) for more efficiently converting between heterogeneous programs for the sake of promoting functional synergy (Safron, 2020a(Safron, , 2020bVanRullen & Kanai, 2021).…”

“…NSCs consist of recurrent neural networks (RNNs) organized into a predictive hierarchy, where each RNN attempts to predict data from the level below, and where only unpredicted information is passed upwards. These predictions are realized in the form of recurrent dynamics (Candadai & Izquierdo, 2020;Lu & Bassett, 2020), whose unfolding entails a kind of search processes via competitive/cooperative attractor formation over states capable of most efficiently responding to-or resonating with (Rumelhart & McClelland, 1987;Safron, 2020b)ascending data streams. In this way, much as in FEP-AI, predictive abilities come nearly "for free" via Hamilton's principle of least action (Ali et al, 2021;K.…”

“…(Reprinted from Safron, 2021b.) Depiction of the human brain in terms of entailed aspects of experience (i.e., phenomenology), as well as computational (or functional), algorithmic, and implementational levels of analysis (1983;Safron, 2020b). A phenomenological level is specified to provide mappings between consciousness and these complementary/supervenient levels of analysis.…”

“…Neuroanatomically speaking, these highly connected areas would correspond to the brain's "rich club" networks (Heuvel et al, 2012), including the 2D grid structures described above (Figure 3), which could contribute to spatiotemporal modeling (Haun & Tononi, 2019) in both concrete physical and abstract (via spatialization) domains. Theoretically, these subnetworks (entailing shared latent space) may be well-modelled as graph neural networks (GNNs) (Safron, 2020b;Zhou et al, 2019), which are gaining increasing popularity as a means of efficiently modelling a broad range of processes. Further, the regulation of neuronal dynamics by diffuse neuromodulator systems could be computationally understood as parameterizing inference and learning with respect to the formation of partially disentangled features in perception, as well as through the selecting and sculpting of particular policies for enaction (e.g.…”