Conditions under which emergence optimization improves transfer learning in reservoir computing

Determine the conditions under which hyperparameter optimization for causal emergence—quantified in this study by positive ψ or high P(E) of the forecast relative to reservoir states—improves transfer learning performance when reservoir computers trained to forecast one chaotic dynamical environment are evaluated on previously unseen environments.

Background

The paper studies reservoir computers trained to forecast environmental dynamics of chaotic systems (Lorenz and Sprott flows) and quantifies emergent dynamics using ψ, a lower bound on causal emergence derived from Partial Information Decomposition. Across tasks and topologies, the authors find a bidirectional coupling between prediction performance and emergence.

To probe generalization, they evolve reservoirs under objectives that weight prediction success P(S) and emergence P(E), then evaluate the fittest models on non-optimized environments. They observe that while optimizing for P(S) tends to yield the best in-domain performance, optimizing for P(E) can sometimes yield competitive or superior performance out-of-domain, motivating an explicit open question about when emergence optimization enhances transfer learning.