Neuroscience · Complexity

Matrix of Consciousness

A thermostat registers heat; a brain registers heat and minds it. Integrated information theory pins the difference on structure: a system is conscious to the degree that its parts carry more information together than apart. Wire four units into a matrix, cut it along its weakest seam, and the bits the whole keeps that the cut cannot are Φ.

Φ
integrated information, in bits: what the whole predicts about its own next step that its best cut does not
EI
effective information: the mutual information between a uniform spread of present states and the noised states that follow
cut
one of the seven ways to split four units in two, with every connection across the split deleted — the weakest cut is the one whose halves keep the most, the seam that costs the system least
ε
noise per unit per step, 6% by default: at zero the future is a lookup table, at 50% it is a coin toss, and either way Φ collapses

Integration lab

Click a cell to cycle it: + excites, − inhibits, · cuts. Rows are the unit that receives and columns the unit that sends, so the diagonal is a unit listening to itself. Φ is rebuilt from the matrix on every change.

Φ = 1.35 bits · cut A, C | B, D
solid excites · dashed inhibits
row listens · column speaks + excite  − inhibit
6%
Integrated information
1.35bits

What noise does to it

The solid curve is the matrix you have wired; the dashed curve is the plain feedforward chain, scored the same way.

this matrix feedforward chain

The horizontal axis is noise per unit per step, on a log scale; the vertical axis is in bits. A loop stays ahead of a chain at every level until both vanish, because a cut through a loop breaks a transition the chain never needed. At 50% noise every present state produces the same futures and the curves arrive at zero together.

The same four units, five architectures

Each row is scored with the same rule, the noise on the slider and the same seven cuts.

Network Links EI whole Φ

What the matrix shows — and what it leaves out

Density is not integration

The all-to-all matrix fires every unit whenever any one of them fires, so sixteen links carry a single bit: cutting it in half predicts as well as the whole. A plain loop scores higher with four links, because each unit’s next state still depends on where the activity is. What a cut destroys is the measure of what the matrix was holding.

Four units, not a brain

Effective information is linear in the number of units, so four bits is the ceiling here and noise takes a bite out of even that. The full theory scores the cause and effect repertoire of every mechanism against every possible cut, which is why practical analyses use a handful of units or approximations rather than a whole cortex. A high Φ on this page is a statement about wiring, not evidence of experience.

The hard problem stays put

IIT’s bold move is to identify experience with the integrated structure itself rather than with something the structure produces. If that is right, Φ does not correlate with feeling, it is what feeling is. Critics answer that no amount of information about structure explains why any of it is like something; in 2023 a letter signed by 124 researchers argued the theory is not even testable. Global workspace, higher-order thought, predictive processing and the attention schema agree on much of the data and disagree about what has to be added.