Performance Prediction Based on Neural Architecture Features

Nerual Architecture Search (NAS) usually requires to train quantities of candidate neural networks on dataset for choosing high performance network architecture and optimizing hyperparameters, which is very time consuming and computationally expensive. In order to resolve the issue we try to use performance prediction method to predict the model performance with little or even no training steps. We assume that the performance is determined once the architecture and other hyperparameters are chosen. So we firstly extract the sequence features of the chain-structured neural architecture by introducing N-grams model to process architecture textual description. Subsequently, based on the extracted neural architecture features, we use appropriate regression model to predict validation accuracies for modeling learning curve. Through a series of experimental comparisons, we verify the effectiveness of our proposed performance prediction method and the effective acceleration of the NAS process.