I’m uncomfortable sending proto-AGI out of distribution. Measuring (conditional) mutual information by how well an auxiliary network predicts some activations from others is potentially asymmetric and intransitive, could that serve?
Yes, the IC and IC* algorithm (described respectively on p. 50 and 52 of Causality) describe how to infer some parts of the causal DAG from a stable probability distribution.
(Not-so-fun fact: Both of these don’t have Wikipedia articles!)
I’m uncomfortable sending proto-AGI out of distribution. Measuring (conditional) mutual information by how well an auxiliary network predicts some activations from others is potentially asymmetric and intransitive, could that serve?
Also, I vaguely remember that Judea Pearl’s Causality explains how to infer causality from mere correlational data.
Yes, the IC and IC* algorithm (described respectively on p. 50 and 52 of Causality) describe how to infer some parts of the causal DAG from a stable probability distribution.
(Not-so-fun fact: Both of these don’t have Wikipedia articles!)