Controlled-Minimax is a new framework that stabilizes adversarial minimax optimization by modeling training dynamics as a second-order ODE and applying a feedback controller derived from control theory. The method guarantees local exponential stability regardless of the Hessian spectrum, addressing persistent limit cycles and divergence that plague traditional gradient descent-ascent approaches. Experiments on benchmarks such as Dirac-GAN, 2D Gaussian mixtures, Swiss Roll, and StyleGAN2-ada show that Controlled-Minimax suppresses oscillations and achieves fast, robust convergence.
arXiv:2410. 07719v4 Announce Type: replace Abstract: Despite being widely adopted as a canonical framework for learning robust models, adversarial training suffers from robust overfitting.
arXiv:2301. 06308v2 Announce Type: replace-cross Abstract: Sharpness-aware minimization (SAM) is a training method that seeks to find flat minima in deep learning, resulting in state-of-the-art performance across various domains.
The paper introduces CoDRA, a cost-to-disturbance ratio approach for adversarial reinforcement learning that balances controller performance and disturbance exposure without extra penalty terms. CoDRA uses a self‑normalized actor–critic update, scaling value terms by a stop‑gradient normalization constant derived from the current batch. Experiments on MuJoCo pendulum tasks show that CoDRA achieves the lowest cost across a range of forces and masses, outperforming other methods especially on the more challenging InvertedDoublePendulum environment.
The paper investigates how deterministic Gradient Descent behaves when applied to Adversarial Training with large, constant step sizes. It analyzes both a linear logistic regression setting—examining where parameters converge (implicit bias) and how fast losses decrease (optimization)—and neural network training, focusing on the evolution of loss sharpness. The study reveals three key distinctions from Standard Training: the implicit bias gap persists even with large steps, there is no uniform convergence rate for optimization error, and sharpness follows a two‑stage dynamic unlike the Edge‑of‑Stability pattern seen in Standard Training.
The paper presents a unified framework for regularization-based robust reinforcement learning by deriving upper bounds on the performance gap between nominal and worst-case policies. These bounds are expressed as a regularization objective plus a KL-divergence penalty, explaining why KL penalties enhance robustness. The authors reformulate robust training as a constrained optimization problem, updating the Lagrange multiplier jointly with the policy to automatically tune regularization, and validate the approach with extensive adversarial evaluations on continuous control tasks.
arXiv:2606. 04476v1 Announce Type: new Abstract: In this paper, we study the gradient descent dynamics for jointly training both layers of a one-hidden-layer ReLU network to fit a linear target function.
The paper introduces a penalized distributionally robust optimization framework that allows an adversary to choose any distribution while incurring a Wasserstein penalty for deviating from the empirical distribution. It shows that the adversary’s problem can be reformulated as optimizing transport maps that push empirical samples to adversarial ones, proving that optimal maps are cyclically monotone. The authors argue that standard per-sample adversarial training violates this property and propose two remedies—multi-start particle ascent and input-convex neural network parameterization—to enforce cyclical monotonicity, demonstrating improved robustness and generalization in experiments on regression, image classification, and control tasks.
The paper analyzes Gradient Descent (GD) on a minimal binary classification problem with a two‑neuron ReLU network trained on two instances. It proves that GD converges to an optimal robustness margin—maximizing the distance between the decision boundary and training points—but does so at a very slow rate, specifically θ(1/ln(t)). Empirical simulations confirm this slow convergence across various initializations, and the authors provide a rigorous analysis of GD trajectories across different activation patterns to establish the first explicit lower bound on robustness convergence in a non‑linear model.
arXiv:2505. 13196v3 Announce Type: replace-cross Abstract: We introduce Velocity-Regularized Adam (VRAdam), a physics-inspired optimizer for training deep neural networks that draws on ideas from quartic terms for kinetic energy with its stabilizing effects on various system dynamics.
The paper introduces StablePDENet, a physics-informed adversarial training framework that regularizes the residual sensitivity of neural operators to improve stability against input perturbations. The method formulates operator learning as a min–max optimization, where a physics-based projected‑gradient adversary generates perturbations and the outer objective combines the attacked physics loss with a normalized residual‑sensitivity penalty. Experiments on benchmark problems show that StablePDENet outperforms PI‑DeepONet and its adversarial variant in accuracy under adversarial attacks while maintaining competitive performance on clean data, and it also enhances generalization and highlights the distinction between model sensitivity and intrinsic operator ill‑conditioning.
The paper derives an exact discrete‑time law that captures how learning‑rate schedules and weight decay interact in scale‑invariant neural networks, showing that a single scalar quantity governs the effective step size. It demonstrates that the balance point between contraction and expansion is intrinsically unstable, leading to recurrent dynamics when using constant learning rates with weight decay. The authors extend this analysis to various optimizers and datasets, confirming the law’s precision and showing that performance peaks sharply at the predicted boundary.
arXiv:2606. 15551v1 Announce Type: new Abstract: The Edge of Stability (EoS) phenomenon, where gradient descent operates with sharpness exceeding the classical convergence threshold yet the loss decreases over long timescales, is ubiquitous in modern deep learning but remains poorly understood in realistic settings.