<ul><li>Performance predictors are essential in neural architecture search (NAS) to speed up the evaluation phase by estimating the performance of new designs based on trained architectures.</li><li>Current predictors often struggle with generalization due to the shift in data distribution between training and testing samples, leading to inaccurate predictions based on spurious correlations.</li><li>To combat this issue, the Causality-guided Architecture Representation Learning (CARL) method has been introduced to differentiate critical and redundant features of architectures for more accurate and interpretable performance prediction.</li><li>Extensive experiments across five NAS search spaces have shown that CARL achieves state-of-the-art accuracy and enhanced interpretability, such as reaching 97.67% top-1 accuracy on CIFAR-10 with DARTS.</li></ul>

CARL: Causality-guided Architecture Representation Learning for an Interpretable Performance Predictor

Discover more