Towards Leveraging Structure for Neural Predictor in NAS

Neural Architecture Search (NAS), which automatically designs a neural architecture for a specific task, has attracted much attention in recent years. Properly defining the search space is a key step in the success of NAS approaches, which allows us to reduce the required time for evaluation. Thus, late strategies for searching a NAS space is to leverage supervised learning models for ranking the potential neural models, i.e., surrogate predictive models. The predictive model takes the specification of an architecture (or its feature representation) and predicts the probable efficiency of the model ahead of training. Therefore, proper representation of a candidate architecture is an important factor for a predictor NAS approach. While several works have been devoted to training a good surrogate model, there exits limited research focusing on learning a good representation for these neural models. To address this problem, we investigate how to learn a representation with both structural and non-structural features of a network. In particular, we propose a tree structured encoding which permits to fully represent both networks’ layers and their intra-connections. The encoding is easily extendable to larger or more complex structures. Extensive experiments on two NAS datasets, NasBench۱۰۱ and NasBench۲۰۱, demonstrate the effectiveness of the proposed method as compared with the state-of-the-art predictors.

کلیدواژه ها:

Neural Architecture Search ، Search Space Pruning ، Network Architecture ، Representation Learning