Source-linked AI summary
Integrated information theory (IIT) 4.0: Formulating the properties of phenomenal existence in physical terms
Larissa Albantakis, Leonardo Barbosa, Graham Findlay, Matteo Grasso, Andrew M Haun, William Marshall, William GP Mayner, Alireza Zaeemzadeh, Melanie Boly, Bjørn E Juel, Shuntaro Sasai, Keiko Fujii, Isaac David, Jeremiah Hendren, Jonathan P Lang, Giulio Tononi
TL;DR
IIT 4.0 addresses how the properties of subjective experience can be explained in operational physical terms. It translates five phenomenal axioms into substrate postulates and a mathematical framework, proposing that an experience is identical to the Φ-structure unfolded from a maximal substrate. The framework supports testable predictions about consciousness and its substrate, while remaining a work in progress whose assumptions and empirical adequacy require further assessment.
Problem
IIT seeks an objective, physical account of subjective experience, including why consciousness depends on particular brain portions and states and why experiences have their specific qualities.
Method
IIT translates five essential phenomenal axioms into necessary and sufficient physical postulates and expresses them mathematically for systems of interacting units.
Results
The framework proposes that each experience corresponds to the cause–effect structure unfolded from a maximal substrate and makes predictions about conscious and unconscious states.
Takeaways & Limitations
IIT’s Φ-structures are intended to account for the quality and quantity of experience and support inferences about consciousness across different systems.
Takeaways & Limitations
IIT remains a work in progress, and its assumptions, axiomatic basis, and relationship between consciousness and its brain substrate remain open to empirical assessment.
Abstract
from arXiv · showhide
This paper presents Integrated Information Theory (IIT) 4.0. IIT aims to account for the properties of experience in physical (operational) terms. It identifies the essential properties of experience (axioms), infers the necessary and sufficient properties that its substrate must satisfy (postulates), and expresses them in mathematical terms. In principle, the postulates can be applied to any system of units in a state to determine whether it is conscious, to what degree, and in what way. IIT offers a parsimonious explanation of empirical evidence, makes testable predictions, and permits inferences and extrapolations. IIT 4.0 incorporates several developments of the past ten years, including a more accurate translation of axioms into postulates and mathematical expressions, the introduction of a unique measure of intrinsic information that is consistent with the postulates, and an explicit assessment of causal relations. By fully unfolding a system's irreducible cause-effect power, the distinctions and relations specified by a substrate can account for the quality of experience.
Introduction
IIT 4.0 aims to explain phenomenal properties in operational physical terms by translating experience’s essential axioms into physical postulates and a mathematical framework. It proposes that an experience is identical to the cause–effect structure unfolded from a maximal substrate.
- Motivation: IIT starts from experience itself and seeks to explain why consciousness depends on particular portions and states of the world.The theory also addresses why experiences have specific phenomenal qualities and makes predictions about their presence and quality.
- Axioms and postulates: Experience is characterized by five essential properties: intrinsicality, information, integration, exclusion, and composition.These correspond to being for the experiencer, specific, unitary, definite, and structured.
- Axioms and postulates: IIT translates the five phenomenal axioms into physical postulates requiring a substrate’s cause–effect power to be intrinsic, specific, unitary, definite, and structured.The translation rests on operational physicalism and related methodological assumptions.
- Core identity: IIT proposes that an experience is identical to the cause–effect structure unfolded from a maximal substrate, with no additional ingredients.The Φ-structure is intended to account for the specific phenomenal properties of experience.
- Formal framework: IIT 4.0 formalizes the postulates mathematically to evaluate consistency, make predictions about experience and its brain substrate, and extrapolate to other beings.The article presents the framework as a reference, a demonstration of internal consistency, and an illustration of computational unfolding.
- Advances in IIT 4.0: IIT 4.0 adds a more accurate axiom-to-postulate translation, the Intrinsic Difference measure, and explicit assessment of causal relations.It is presented as more complete and self-consistent than earlier IIT formulations.
From phenomenal axioms to physical postulates
IIT begins with phenomenal existence and identifies five essential properties of experience, then infers corresponding physical requirements for a conscious substrate. These postulates are presented as necessary and sufficient, culminating in a Φ-structure that accounts for experience’s quality.
- Phenomenal axioms: IIT treats the existence of experience as immediate and irrefutable, making it the starting point for identifying phenomenal axioms.The theory calls this existence the ultimate axiom and basis for further inferences.
- Phenomenal axioms: The five axioms state that experience is intrinsic, specific, unitary, definite, and structured.They correspond respectively to existing for itself, being the way it is, forming an irreducible whole, being this whole, and containing distinctions and relations.
- Scope of the axioms: IIT excludes space, time, change, self, intentionality, affect, and other candidates from axiomatic status because experiences lacking each remain conceivable and achievable.The paper points to dreaming, meditation, and drugs as examples of altered states supporting this scope claim.
- Methodological commitments: IIT’s realism treats the world as persisting independently of experience while using observations and manipulations to characterize physical existence operationally.The framework presents this as a better hypothesis than solipsism because it explains and predicts regularities of experience.
- Physical postulates: Physical existence is defined operationally as cause–effect power: a substrate must be able to take and make a difference.The substrate is understood as a set of observable and manipulable units, consistent with operational physicalism and atomism.
- Physical postulates: The postulates require intrinsic, specific, irreducible, definite, and structured cause–effect power corresponding to the five phenomenal axioms.Specificity selects a maximal intrinsic-information state; integration measures irreducibility with integrated information; exclusion selects the maximally irreducible unit set.
- Complexes and Φ-structures: A maximal substrate, or complex, is the unit set with maximal integrated information, whose subsets specify overlapping distinctions and relations forming a Φ-structure.The Φ-structure corresponds to experience quality, while the associated integrated-information values quantify its structure’s amount.
- Explanatory identity: IIT claims the postulate-defined physical properties are necessary and sufficient for consciousness and that the unfolded Φ-structure fully explains experience without ad hoc ingredients.This is the theory’s explanatory identity linking phenomenal properties to physical properties.
Overview of IIT’s framework
IIT 4.0 analyzes a substrate from its transition probability matrix, first identifying a maximal substrate and then unfolding its Φ-structure from distinctions and relations. The framework evaluates cause–effect power through intrinsic information and integrated information, while acknowledging computational and modeling limits.
- Substrate representation: IIT characterizes a substrate by the transition probability function of its interacting units and applies the postulates to identify candidate conscious substrates.The substrate is modeled as a stochastic system with finite state space, discrete updates, and conditionally independent unit updates.
- Substrate selection: For each candidate system, IIT selects a maximal cause–effect state using intrinsic information and then selects a maximal substrate using integrated information.The winning substrate must maximize integrated information relative to overlapping competitors.
- Φ-structure: The framework unfolds a complex into distinctions specified by mechanisms over purviews and relations formed by congruent overlaps among distinctions.The resulting Φ-structure corresponds to experience quality, while the sum of distinction and relation values corresponds to quantity.
- Substrate representation: The transition probability matrix describes each possible state transition under interventions that initialize the system in every possible state.It serves as the overall description of the system’s cause–effect power.
- Postulate assessment: IIT evaluates whether a system satisfies intrinsic, specific, integrated, definite, and structured cause–effect requirements.These requirements are assessed from a complete interventional transition-probability description.
- Information measures: Intrinsic information measures causal informativeness for a specified cause and effect state, while selectivity measures concentration over that state relative to alternatives.The product of informativeness and selectivity captures a tension between expansion and dilution as systems include more units.
- Integration: Integrated information quantifies how much intrinsic information is lost under a partition, thereby measuring irreducibility of the maximal cause–effect state.The measure is sensitive to fault lines separating weakly or directionally interacting parts.
- Exclusion: Exclusion selects a definite set of units, grain, updates, and states by maximizing integrated information over competing overlapping systems.The selected set is called a maximal substrate or complex, and complexes can be identified recursively in a universal substrate.
Box 2. Ontological principles of IIT
IIT’s ontological principles define existence through cause–effect power, selecting maximally irreducible complexes and minimally irreducible partitions. The resulting complex is unfolded into causal distinctions and relations that constitute its Φ-structure and correspond to experience.
- Being: Existence requires cause–effect power: a system exists physically insofar as it can take and make a difference.IIT operationalizes being as the capacity to bear a cause and produce an effect.
- Maximal existence: Maximal existence selects the overlapping candidate complex with the greatest system integrated information, ϕs.Lower-ϕs candidates overlapping the same substrate are excluded from existence.
- Minimal existence: Minimal existence evaluates irreducibility across the partition where the system is least irreducible, the minimum partition.The same principle applies to distinctions, which must specify both an irreducible cause and an irreducible effect.
- Intrinsic information: Intrinsic information measures how a system in its current state selects a specific cause or effect state over itself.The maximally informative cause–effect state is identified before integrated information assesses its irreducibility.
- Integrated information: System integrated information, ϕs, quantifies irreducible cause–effect power over itself as one system and thereby measures irreducible existence.A recursive search identifies maximal substrates, or complexes, among competing candidate systems.
- Φ-structure: A maximal substrate satisfies intrinsic, specific, irreducible, definite, and structured postulates, existing intrinsically as an unfolded set of causal powers.Its Φ-structure corresponds to experience’s quality, while its Φ value corresponds to its quantity.
Results and discussion
The examples show that connectivity, integration, specialization, and unit state determine which substrates qualify as complexes and how richly they specify Φ-structures. They also show that small state changes and functional equivalence can yield substantially different intrinsic cause–effect structures.
- Connectivity and complexes: Bottleneck architectures with high degeneracy and indeterminism fragment into small complexes, limiting the richness of their Φ-structures.In Fig. 6A, units Ab form one complex while c, d, e, and f form monads.
- Connectivity and complexes: Weakly interconnected modules form separate small complexes, while modular or physiological fault lines reduce integrated information.The example connects architectural modularity and sleep-related causal ineffectiveness with reduced consciousness.
- Connectivity and complexes: A six-unit directed cycle forms a large complex with ϕs = 1.74 ibits but low structured information, Φ = 7.65, because it lacks higher-order cross-connections.Highly deterministic cycles can be large and irreducible yet specify sparse Φ-structures.
- Connectivity and complexes: Specialized, effectively connected lattices support richer Φ-structures, with an optimally connected n-unit system bounded by 2n −1 distinctions.The six-unit specialized lattice is maximal and contains 24 distinctions out of 63 possible distinctions.
- Connectivity and complexes: Connectivity strongly determines which units form complexes and the structured information their cause–effect structures specify.The examples contrast bottleneck, modular, cyclic, specialized, and partially integrated architectures.
- State and activity: Changing one unit from active to inactive can alter purviews, irreducibility, relations, and Φ, whereas permanently inactivating it collapses its causal contribution and shrinks the complex.The five-unit example changes from 21 distinctions and 4760 relations at ABcdE to a four-unit complex with 14 distinctions and Φ = 3.31 after inactivation.
A.1 Ties
IIT addresses ties among systems, states, and mechanisms by applying postulate-based selection rules that preserve maximal intrinsic cause–effect power and structured information.
- A.1 Ties: Symmetries can produce ties among multiple unit sets or cause–effect states for maximal intrinsic information or cause–effect power.IIT explicitly outlines tie-resolution procedures to remain consistent with its postulates and principles.
- A.1 Ties: For systems, an iterative algorithm identifies maximal substrates and excludes overlapping systems with lower ϕs from existing as complexes.When overlapping systems tie for maximality, the procedure selects the next best unique system.
- A.1 Ties: A system state is selected by maximal intrinsic cause-and-effect information, with remaining ties resolved by maximizing integrated information ϕs over the minimum partition.Residual ties in highly symmetrical systems can specify intrinsically identical cause–effect structures and therefore remain extrinsic.
- A.1 Ties: A mechanism’s maximally irreducible cause or effect is selected by comparing integrated information ϕ across maximal states and all possible purviews.The maximum existence postulate governs potential ties in the resulting cause–effect state.
- A.1 Ties: Within a purview, tied states are resolved by selecting the state congruent with the system’s cause–effect state s′.Across purviews, the maximal state generally favors the one supporting the most relations, typically through larger purviews.
A.2 Comparison to IIT 1.0–3.0 and subsequent publications
IIT 4.0 presents a more complete and self-consistent formulation while retaining the theory’s core framework. Its advances include progressively refined formalizations and a new Intrinsic Difference measure.
- IIT 4.0 is described as a more complete, self-consistent formulation than IIT 1.0, 2.0, and 3.0.
- The theory’s core has remained stable while its formal framework has been progressively refined and extended.
- A notable advance is introducing an Intrinsic Difference measure uniquely consistent with IIT’s postulates.
Axioms and postulates
IIT 4.0 distinguishes phenomenal existence from its essential properties and refines the axioms and postulates to align more closely with one another.
- Axioms and postulates: IIT 4.0 treats phenomenal existence as a foundational zeroth axiom and identifies intrinsicality, information, integration, exclusion, and composition as five essential properties.These five axioms are presented as immediate and irrefutably true of every conceivable experience.
- Axioms and postulates: The updated exposition separates phenomenal existence, which is not itself a property, from intrinsicality, which is an essential property of phenomenal existence.This distinction clarifies the conceptual organization of IIT’s axiomatic basis.
- Axioms and postulates: Compared with IIT 3.0, IIT 4.0 refines the axioms and updates the postulates to track the phenomenal axioms more closely.For example, the information postulate requires selection of a specific cause–effect state over the system’s units.
Identifying maximal substrates
IIT identifies maximal substrates by first selecting a system’s maximal cause–effect state and then evaluating its integrated information.
- Identifying maximal substrates: Maximal substrates are identified by evaluating system integrated information ϕs for the maximal cause–effect state.IIT 4.0 identifies the minimum partition as the partition with minimal normalized integrated information across the maximal possible value.
Measuring intrinsic information
IIT 4.0 replaces earlier divergence-based measures with Intrinsic Difference, a measure designed to comply uniquely with IIT’s postulates. Intrinsic information is evaluated from informativeness and selectivity.
- IIT 4.0 replaces Kullback-Leibler divergence and extended Earth Mover’s Distance with the Intrinsic Difference measure.
- Intrinsic Difference is presented as uniquely complying with IIT’s postulates.
- Intrinsic information is evaluated as the product of informativeness and selectivity.
Causal distinctions
IIT 4.0 reformulates causal distinctions as irreducible specifications of specific cause and effect states. The updated procedure selects these states using Intrinsic Difference and applies revised partition constraints.
- Causal distinctions: Causal distinctions require irreducibility, expressed as ϕd > 0, and select a specific cause state and effect state.
- Causal distinctions: A mechanism selects its specific cause and effect state using the Intrinsic Difference measure.
- Causal distinctions: The updated formulation selects the cause–effect state before evaluating integrated information, so selection depends on the mechanism as a whole rather than its partitions.
- Causal distinctions: A candidate distinction contributes to the system’s cause–effect structure only when its maximal state is congruent with the system’s maximal state.
Relations
IIT 4.0 makes relations explicit as causal bindings among distinctions with congruent overlaps. It also provides analytical procedures for computing relation information and treating self-relations without combinatorial explosion.
- Relations: Relations bind causal distinctions over congruent overlaps between their cause and/or effect purviews.
- Relations: A system is a conscious complex when it corresponds to a maximum of system integrated information determined through information, integration, and exclusion.
- Relations: The experience’s quality corresponds to the Φ-structure of distinctions and relations, while its quantity corresponds to the Φ value.
- Relations: Any subset of at least two distinctions sharing a node defines a relation whose overlap contains that node.
- Relations: The sum of relation integrated information can be reorganized over nodes and relations whose purviews contain those nodes, with overcounting corrected by joint purview size.
- Relations: Self-relations can be assessed individually without combinatorial explosion after analytically summing non-self-relations.
Analytical count of the number of relations
The paper counts causal relations among distinctions by generalizing subset-based definitions and applying inclusion-exclusion, while treating self-relations separately to avoid combinatorial explosion.
- Causal relations are counted among all distinctions in D(TS, s).
- The counting procedure generalizes Z(n) to all subsets o ⊆s′.
- Each distinction d ∈D(TS, s) corresponds to an element e(d), and subsets of Z(o) with at least two elements define relations whose overlap contains o.
- Inclusion-exclusion is applied to count the relations, excluding self-relations from the resulting total.
- Self-relations are counted individually without combinatorial explosion.