Efficient and Lightweight Differentiable Architecture Search

While Neural Architecture Search (NAS) has revolutionized the automation of deep learning model design, gradient-based approaches like DARTS often suffer from high computational overheads, the collapse of skip-connections, and optimization instability. To address these limitations, we propose Efficient and Lightweight Differentiable Architecture Search (EL-DARTS). EL-DARTS constructs a compact and redundancy-reduced search space, integrates a partial channel strategy to lower memory usage, employs a Dynamic Coefficient Scheduling Strategy to balance edge importance, and introduces entropy regularization to sharpen operator selection. Experiments on CIFAR-10 and ImageNet demonstrate that EL-DARTS substantially improves both search efficiency and accuracy. Remarkably, it attains a 2.47% error rate on CIFAR-10, requiring merely 0.075 GPU-days for the search. On ImageNet, the discovered architecture achieves a 26.2% top-1 error while strictly adhering to the mobile setting (<600 M MACs). These findings confirm that EL-DARTS effectively stabilizes the search process and pushes the efficiency frontier of differentiable NAS.