Neural architecture search (NAS) is a core technology in AutoML, but it faces challenges such as high computational costs and inefficient evaluation. Traditional NAS methods require fully training each candidate architecture, leading to substantial resource consumption and long evaluation times. This paper introduces a four-stage progressive multi-task learning framework that shifts from brute-force search to performance prediction. By progressively training from simple synthetic data to complex real data, the framework enables efficient architecture performance prediction. The main contributions are a unified progressive predictor paradigm, a deployment-aware multi-task prediction mechanism with dual lightweighting, and a benchmark-aware data-and-transfer framework based on NATS-Bench reconstruction and progressive knowledge distillation. Experiments on 15,000 NATS-Bench architectures with a fixed train–validation–test split (8:1:1) and consistent hyperparameters show a 95.56% correlation-based prediction score, computed as Pearson correlation expressed as a percentage (Pearson correlation 0.9556, R-squared 0.9134), 32-fold training efficiency improvement (37.87 s ± 2.1 s vs. 1247.6 s), and 95.7% convergence stability across five random seeds. Ablation studies, including a Direct Stage-3-Only comparison, quantify component contributions, and benchmarks compare against random forest, XGBoost, and MLP under identical data splits and feature spaces.
Wang et al. (Tue,) studied this question.