Deep learning models have achieved striking performance across vision, language and time-series tasks, yet their growing depth and parameter counts impose substantial computational and memory demands.