Minimum Fidelity for Reliable Architecture Ranking in Bayesian NAS for Object Detection

This paper addresses the problem of reducing computational costs in Neural Architecture Search for object detection. The key question in low-fidelity approaches is: after what minimum number of epochs does the early training signal already provide acceptable architecture ranking? This paper presents a methodology for determining minimum fidelity based on out-of-sample rank correlation: trials are split into calibration (70%) and test (30%) sets, a composite proxy is built on the first set and evaluated on the second. An empirical study was conducted on 60 architectures trained for 25 epochs on an object detection dataset (6,772 images, 6 classes). The results show that a rank ensemble of three training metrics (val_loss, train_accuracy, val_accuracy) achieves Spearman ρ = 0.877 out-of-sample at epoch 3, providing 88% computational savings. A more complex 7-component metric underperforms the simple ensemble due to overfitting on a small sample (ρ_test = 0.70 vs. 0.88). The results are locally valid within the studied search space and the specific dataset.