Latent Space Neural Architecture Search via LambdaNDCGloss-Based Listwise Ranker

The rapid development of neural architecture search (NAS) promotes the research of efficient evaluation of candidate architectures. However, as one of highly efficient evaluation methods, the predictor is difficult to accurately predicting the absolute performance of architecture. Moreover, it is difficult for a predictor to obtain the high prediction performance in discrete architecture representation space. Thus, a LambdaNDCGloss-based listwise ranker is developed to accurately predict the architecture through ranking prediction. The LambdaNDCGloss assigns large weights to architectures with real top ranking during the training process. Then, architectures with discrete representation are embedded into the latent continuous space through graph variational autoencoder to further enhance the prediction performance of ranker. Moreover, a training data reconstruction method is proposed to divide the entire training set into several subsets, from where the ranker can learn the ranking distribution more easily in contrast to that of the entire training set. Extensive experiments on NAS benchmarks demonstrate that our approach outperforms the state-of-the-art methods.