Autonomously and Simultaneously Refining Deep Neural Network Parameters by Generative Adversarial Networks

  • 2018-05-24 15:00:13
  • Burak Kakillioglu, Yantao Lu, Senem Velipasalar
  • 5

Abstract

The choice of parameters, and the design of the network architecture areimportant factors affecting the performance of deep neural networks. However,there has not been much work on developing an established and systematic way ofbuilding the structure and choosing the parameters of a neural network, andthis task heavily depends on trial and error and empirical results. Consideringthat there are many design and parameter choices, such as the number of neuronsin each layer, the type of activation function, the choice of using drop out ornot, it is very hard to cover every configuration, and find the optimalstructure. In this paper, we propose a novel and systematic method thatautonomously and simultaneously optimizes multiple parameters of any given deepneural network by using a generative adversarial network (GAN). In our proposedapproach, two different models compete and improve each other progressivelywith a GAN-based strategy. Our proposed approach can be used to autonomouslyrefine the parameters, and improve the accuracy of different deep neuralnetwork architectures. Without loss of generality, the proposed method has beentested with three different neural network architectures, and three verydifferent datasets and applications. The results show that the presentedapproach can simultaneously and successfully optimize multiple neural networkparameters, and achieve increased accuracy in all three scenarios.

 

Quick Read (beta)

loading the full paper ...