sharpness-aware

Definition

A training objective that penalizes not just the loss at the current parameters but how steeply loss rises around them. The best-known instance, Sharpness-Aware Minimization, first steps toward the worst-case nearby perturbation, then descends the loss measured there — effectively steering optimization into wide, flat basins. It improves generalization on standard vision benchmarks; the cost is roughly double the gradient computation per update.