Vanilla tanh RNN (task-trained, N=100 neurons) — Math in LLMs Research Map

MATH · IN · MODELS

Vanilla tanh RNN (task-trained, N=100 neurons)

various (academic, computational-neuroscience tradition)

Structures found in this family (1)

Ghost Point (saddle-node bifurcation remnant)

By model (2)

Full-rank RNN (N=100, working-memory tasks, Dinc et al. 2025) · 100 units, full-rank

Ghost Point (saddle-node bifurcation remnant)

Rank-one RNN (N=100, working-memory tasks, Dinc et al. 2025) · 100 units, rank-1

Ghost Point (saddle-node bifurcation remnant)

Observations (1)

Ghost points of destroyed saddle-nodes govern abrupt RNN learning

Papers

A Ghost Mechanism: An Analytical Model of Abrupt Learning in Recurrent Networks (2025)