Popular repositories Loading

1. negative-drift-scaling negative-drift-scaling Public

   Interactive demo, scripts, and paper for “Anti-Dissipative Dynamics in Neural Network Optimization” exploring negative drift scaling in symplectic optimization.

   Python

2. geometric-training-invariants geometric-training-invariants Public

   Geometric invariant probes that certify quasi-symplectic structure in neural network training. Reveals that BatchNorm breaks the underlying geometry while GroupNorm preserves it.

   Python

3. tiny-worlds tiny-worlds Public

   Minimal RNNs learn world-aligned dynamics at H=2-32. GRUs develop high-Q resonant manifolds (λ₁≈0); RNNs are forced oscillators. Includes null-world hallucination tests and Lyapunov analysis.

   Python