A novel HashedNets model based on the efficient hyperparameter optimization

Abstract

The research of neural networks compression becomes a hot spot in the AI area. In this paper, we propose a novel method to optimize the hyperparameters of a compression Neural Networks called HashedNets with Radial Basis Function (RBF) interpolation model and Dynamic Coordinate Search (DYCORS) method, the proposed model is called HD-HORD which can help the HashedNets search for the best hyperparameters automatically and efficiently. Experimental results show that the efficiency of HD-HORD can be improved 72% faster than other methods.