Tag
The paper proposes GEAR-SAM, which adaptively allocates the perturbation budget across network blocks using an exponential moving average of squared gradients, improving generalization without additional computational overhead.