Controlled-Minimax: Stabilizing Adversarial Optimization via Control Theory
Controlled-Minimax treats adversarial training as a physical system governed by second-order differential equations to diagnose why optimization fails near equilibrium points. It introduces a specific feedback controller designed to dissipate system energy, guaranteeing local exponential stability regardless of the loss function's Hessian structure. This provides a rigorous method for stabilizing complex minimax problems.