Effect Verification of a Feature Extraction Method Based on Graph Convolutional Networks

Designing effective and efficient neural network architecture is critical to the success of Neural Architecture Search (NAS). One of the critical issues we encounter in Neural Architecture Search is the neural architecture performance evaluation, and a vital step during the search progress is the neural architecture's feature extraction method. The feature vector from the input neural network was mapped to the latent space in the previous work. This feature vector is used to predict the performance of the neural architecture through a trained accuracy decoder. Previous research has employed vector-based ways to predict the neural architecture performance. The encoding scheme converts the neural architecture into an encoded vector, then forward the vector into the model for training to obtain the final prediction result. Therefore, this method tends to ignore the spatial information contained in the structure. According to specific research, each neural architecture may be treated as a graph-based data structure, allowing us to use Graph Convolutional Networks (GCN) to extract features from neural architectures. The neural architecture does not need to be converted into vectors in GCN. Instead of losing structural space information during vector translation, we can extract features straightly from the graph. Experimental results on the NASBench-101 dataset show that GCN methods can predict the accuracy of neural architectures better more accurately than state-of-the-art methods, and our method can evaluate the performance of specific neural architectures effectively and efficiently.