Semi-supervised Regression with Generative Adversarial Networks Using Minimal Labeled Data
Greg Olmschenk · CUNY Academic Works (City University of New York) · 2019
This work studies the generalization of semi-supervised generative adversarial networks (GANs) to regression tasks. A novel feature layer contrasting optimization function, in conjunction with a feature matching optimization, allows the adversarial network to learn from unannotated data and thereby reduce the number of labels required to train a predictive network. An analysis of simulated training conditions is performed to explore the capabilities and limitations of the method. In concert with the semi-supervised regression GANs, an improved label topology and upsampling technique for multi-target regression tasks are shown to reduce data requirements. Improvements are demonstrated on a wide variety of vision tasks, including dense crowd counting, age estimation, and automotive steering angle prediction. With training data limitations arguably being the most restrictive component of deep learning, methods which reduce data requirements hold immense value. The methods proposed here are general-purpose and can be incorporated into existing network architectures with little or no modifications to the existing structure.