Finite experiments certify integrated information in 4-5 unit networks
Original title: Same Behavior, Different Mechanism: What Finite Experiments Can Certify About Integrated Information
Recurrent and feed-forward networks with identical behavior carry different φ values and are separable by internal recordings.
The paper asks when integrated information φ is identifiable from finite partial noisy interventions in stochastic networks of four and five units. Identical input-output laws yield 0.28 and 0 bits yet internal recordings separate the networks within 4-53 trials. Partial access imposes precision floors of 0.30-1.02 bits and the measure has sharp edges with 25-100% changes across edges.
- Identical behavior networks separated by internal recordings in 4-53 trials
- Hidden unit recordings impose precision floors of 0.30-1.02 bits
- Infinitesimal mechanism changes alter φ by 25-100% at most edges
- Lower bounds certified from 10^7 transitions with low false-certificate probability
From the original article
1 min read at arxiv.orgInterpretability, AI auditing and assessments of machine consciousness increasingly rest on mechanism-level quantities: functionals of a system's causal transition law that describe how it computes, not what it outputs. Experiments see that law only through finite, partial and noisy interventions.
By Linqian Wu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.