Generative Adversarial Networks as Weight Generators in Ensemble Classification

Ensemble classifiers are constructed from various component classifiers by aggregating the results of the individual classifiers in some manner. While there are many ways to do this, one of the least computationally intensive is to use a voting ensemble, where the vote of each classifier is weighted by a number representing its performance. Choosing the appropriate weights is an involved process, that ranges from using trial and error to using sophisticated machine learning algorithms. In this paper, we proposed a novel technique that uses generative adversarial networks to customize the weights used for each individual input to the ensemble. Specifically, we propose to use conditional generative adversarial networks to generate the weights, with the input image acting as the label to the network and the appropriate weight being the output. We test the proposed system on an image classification task and show that the proposed method produces better results than using fixed weights by 2.43 percent.