Evaluating Approximations and Heuristic Measures of Integrated Information
Published in Entropy, 2019
Integrated Information Theory (IIT) proposes Φ (Phi) as a formal measure of consciousness grounded in causal structure. Unfortunately, computing true Φ is computationally intractable for all but the smallest systems, motivating a range of approximations and heuristic alternatives used in empirical research.
This paper evaluates 14 such measures—including several widely used in consciousness science—against ground-truth Φ computed for over 2,000 randomly generated small networks (3–6 binary nodes). The key findings are: (1) computational approximations achieve excellent correlation with true Φ (r > 0.95) but provide minimal runtime savings; (2) heuristic measures show reasonable rank-order correlation with state-independent peak Φ, but fail to capture the state-dependent variation that IIT actually predicts to correspond to conscious experience.
The authors conclude that any empirical measure claiming to test IIT must first demonstrate fidelity to true Φ in tractable systems, a criterion most popular heuristics do not meet.
Recommended citation: Sevenius Nilsen A, Juel BE, Marshall W. (2019). "Evaluating Approximations and Heuristic Measures of Integrated Information." Entropy. 21(5):525. https://doi.org/10.3390/e21050525
Download Paper