Propose an architectural modification to the existing neural network to improve its robustness against noisy inputs, and discuss the tradeoff between regularization term strength and model generalization ability.

Implement weight noise with a small value of η to add a regularization term to the objective function. This will help prevent overfitting to noisy data while maintaining performance on clean data. The optimal value of η should be chosen based on the tradeoff between regularization strength and model generalization ability, taking into account the specific characteristics of the problem and the available computational resources.