EOC analysis for biased networks
Background: In the study of dynamical systems, phase transitions are characterized by the stability of a trivial fixed point. This approach has been applied to analyze the edge-of-chaos (EOC) in unbiased recurrent neural networks, but the introduction of bias shifts fixed points and alters the dynamical landscape.
Question / Future Work: Extending the linearization-based EOC analysis from the zero-bias setting to the biased setting remains a significant theoretical challenge, as the origin is no longer a global attractor in biased networks.
Why It Matters: Extending the EOC analysis to biased networks is critical for creating initialization schemes that are robust to realistic weight and bias distributions.
Evidence: we leave the extension to the biased setting s_c > 0, where q_∞ no longer vanishes, to future work.
Metadata & Links
- created_at
- 2026-05-14T06:10:53Z