The Effects of Dropout and Weight Initialization on Human Face Classification Accuracy Using Multiple-agent Generative Adversarial Network

In real-life scenarios, identifying faces is one of the most challenging tasks due to imbalanced data, either binary or multi-class, which compromises the accuracy of many face recognitions methods. In recent years, deep generative models such as Multiple-agent generative adversarial network (MA-GAN) have attracted considerable attention for its superior ability to generate data. MA-GAN outperformed many traditional GAN methods in terms of providing better facial texture details because of its unique model architecture. However, dealing with complex classification situations raises concerns that co-adaptation among neural structures can result in overfitting, which makes it more difficult to provide better facial samples and reduces training time, leading to lower performance. The purpose of this paper is to investigate and test two fundamental concepts as methods of overcoming the limitations of MA-GAN, the dropout and the weight initialization. Currently, no comprehensive study examines data input, network configuration, and learning efficiency simultaneously. Thus, we focus on overfitting and make use of dropouts and weights to increase classification accuracy. We evaluated our method on two datasets: Multi-PIE and CAS-PEAL. Both quantitatively and qualitatively, we show that such methods are capable of optimizing MA-GAN performance.