NASGEM: Neural Architecture Search via Graph Embedding Method

Neural Architecture Search (NAS) automates and prospers the design of neural networks. Estimator-based NAS has been proposed recently to model relationship between architectures their performance enable scalable flexible search. However, existing estimator-based methods encode architecture into a latent space without considering graph similarity. Ignoring similarity in node-based search may induce large inconsistency similar graphs distance continuous encoding space, leading inaccurate representation and/or reduced capacity that can yield sub-optimal results. To preserve correlation information encoding, we propose NASGEM which stands for via Graph Embedding Method. is driven by novel embedding method equipped with measures capture topology information. By precisely estimating using an auxiliary Weisfeiler-Lehman kernel guide utilize additional structural get more accurate improve efficiency. GEMNet, set networks discovered NASGEM, consistently outperforms crafted classification tasks, i.e., 0.4%-3.6% higher accuracy while having 11%- 21% fewer Multiply-Accumulates. We further transfer GEMNet COCO object detection. In both one-stage twostage detectors, our surpasses its manually-crafted automatically-searched counterparts.