Research questionHow can neural networks retain capacity while fitting within fixed parameter and memory budgets?Large models consume substantial memory through their parameters, restricting the depth and width that can be trained or deployed within a given budget. Reducing this cost must not undermine optimization stability or predictive performance.