OP here. This is an experiment to replace standard RNN/Transformer attention mechanisms with a geometric control-theory approach.
The core idea is a "Pilot" (pointer) that physically navigates a 1D Riemann Helix based on gradient flux. We treat learning as a physics problem involving Inertia, Friction (Deadzone), and Stochastic Walk.
It behaves like a quantum system where the particle location (training state) and the wave function (inference state) have de-synced.
The code is raw, manual CUDA. Looking for feedback on the inertia logic and if anyone has seen this specific "Shell vs. Core" divergence before.
OP here. This is an experiment to replace standard RNN/Transformer attention mechanisms with a geometric control-theory approach.
The core idea is a "Pilot" (pointer) that physically navigates a 1D Riemann Helix based on gradient flux. We treat learning as a physics problem involving Inertia, Friction (Deadzone), and Stochastic Walk.
It behaves like a quantum system where the particle location (training state) and the wave function (inference state) have de-synced.
The code is raw, manual CUDA. Looking for feedback on the inertia logic and if anyone has seen this specific "Shell vs. Core" divergence before.