Jiang Zhang

Authored papers (1)

• Emergence and Causality in Complex Systems: A Survey on Causal Emergence and Related Quantitative Studies2023ⓒ 4

  Effective information (EI) — defined as mutual information between a uniformly intervened input distribution and the resulting output distribution of a Markov transition probability matrix — can increase under coarse-graining, a phenomenon Hoel et al. (2013, PNAS 110:19790) termed causal emergence (CE), and this review argues that CE is neither measure-specific nor framework-specific: Comolatti and Hoel showed it appears across Hume's constant conjunction, Cheng's causal attribution, Eells's probability-raising, and Pearl's causal measures alike. The central methodological contribution surveyed is the Neural Information Squeezer Plus (NIS+), an encoder–dynamics-learner–decoder neural architecture that directly maximizes EI using inverse probability weighting and kernel density estimation, and which outperforms the original NIS, variational autoencoders, and feed-forward networks on out-of-distribution data generated by SIR dynamics, the Boid flocking model, and Conway's Game of Life. Applied to two real fMRI datasets — AOMIC ID1000 (movie-viewing) and AOMIC PIOP2 (resting-state, 50 subjects) — NIS+ compressed 100-dimensional micro-state signals to a 1-dimensional macro-state for the task-engaged condition but required 7 dimensions for resting-state, with attribution analysis localizing high-weight micro-signals to visual cortex in the task condition. The paper argues this body of work implies that CE is a principled, data-driven criterion for identifying genuinely causal macro-level representations, that EI maximization suppresses spurious correlations in exactly the way required for out-of-distribution generalization, and that causal emergence identification and causal representation learning are structurally isomorphic problems whose techniques should cross-pollinate.

More papers — OpenAlex / S2

Recent mentions (1)

• papers-typed

  yuan-2023-emergence.md