Pooling Graph Convolutional Networks for Structural Performance Prediction

Neural Architecture Search can help in finding high-performance task specific neural network architectures. However, the training of architectures that is required for fitness computation can be prohibitively expensive. Employing surrogate models as performance predictors can reduce or remove the need for these costly evaluations. We present a deep graph learning approach that achieves state-of-the-art performance in multiple NAS performance prediction benchmarks. In contrast to other methods, this model is purely supervised, which can be a methodologic advantage, as it does not rely on unlabeled instances sampled from complex search spaces.

Automated summary

Neural Architecture Search can find high-performance task specific neural network architectures. Using surrogate models as performance predictors can reduce the need for costly evaluations. Our deep graph learning approach achieves state-of-the-art performance in multiple NAS performance prediction benchmarks.