Neural Architecture Search via Two Constant Shared Weights Initialisations

원문 초록

In recent years, zero-cost metrics are gaining ground in neural architecture search (NAS). There metrics allow finding the optimal neural network for a given task faster and with a lesser computational load than conventional NAS methods. Equally important is that they also shed some light on the internal workings of neural architectures. This paper presents a zero-cost metric that highly correlated with the train set accuracy across the NAS-Bench-101, NAS-Bench-201 and NAS-Bench-NLP benchmark datasets. We evaluate a neural achitecture's potential based on the outputs' statistics after two constant shared weights initialisations. For this, we only use an unlabelled mini-batch of data. We observe that the dispersion of the outputs between two initialisations positively correlates with trained accuracy. The correlation further improves when we normalise dispersion by average output magnitude. The resulting metric, epsilon, does not require gradients computation and unbinds the NAS procedure from training hyperparameters, loss metrics and human-labelled data. Our method is easy to integrate within existing NAS algorithms and takes a fraction of a second to evaluate a single network. The code supporting this study can be found on GitHub at https://github.com/egracheva/epsinas.