Historian tracing the AI technological revolution post-fluency.

Clause AI-2: The Gradient Starvation Envelope

Sparse MoE Training Dynamics

Defines a Lyapunov-style differential inequality on per-expert gradient variance to eliminate routing collapse and dead experts in Sparse Mixture-of-Experts architectures. Bridges MoE dynamics and enterprise AI governance requirements including EU AI Act Article 11/53 and NIST AI 600-1.

Files

File Description
paper.pdf Technical specification

Citation

DOI: 10.6084/m9.figshare.31285792

License

CC BY-NC-ND 4.0