Source-linked AI summary
Basin Geometry and Reliable Recall of Dynamical Memories in Reservoir Computing
Ling-Wei Kong, Ying-Cheng Lai
TL;DR
The paper asks how reliable dynamical-memory recall can persist when autonomous basins are riddled-like and unpredictable at finite precision. It analyzes basin geometry and cue-driven generalized synchronization, finding that recall succeeds when contraction brings the state into a robust basin head, with recall duration linked to synchronization rate and head radius.
Problem
Reliable recall is conventionally associated with broad, coherent attractor basins, but these reservoirs can recall memories despite tentacular regions with near-zero uncertainty exponents.
Method
The paper characterizes high-dimensional basin geometry using slices and perturbation tests, then relates cue-driven contraction to robust-head size through independently measured synchronization rates and recall thresholds.
Results
Recall becomes reliable once generalized synchronization contracts the driven state into the target basin head, and the resulting geometry–dynamics prediction closely matches measured recall times.
Takeaways & Limitations
Robust accessibility at finite precision depends on basin-head radius rather than total basin volume, while similar geometry in a trained RNN suggests the phenomenon is not limited to fixed-matrix reservoirs.
Abstract
from arXiv · showhide
Reliable attractor recall conventionally requires broad basins of attraction. However, in reservoir-computing based associative memory, temporal cues reliably recover dynamical memories despite basins dominated by unpredictable, riddled-like regions. We reveal that memory basins exhibit an ``octopus-like'' structure: a robust ``head'' near the attractor and thin, intertwined ``tentacles'' spanning state space. Initial states in tentacular regions yield near-zero uncertainty exponents, making the recalled memory effectively unpredictable at finite precision. Yet, cue-driven generalized synchronization bypasses this unpredictability, driving the system into the robust basin head. This mechanism yields a quantitative relation linking minimum cue duration, synchronization rate, and basin-head radius. Trained recurrent neural networks exhibit similar geometry, suggesting this phenomenon extends beyond reservoir computing.