Robust Semi-Supervised Classification using GANs with Self-Organizing Maps

TL;DR

This work describes an architecture that combines self-organizing maps (SOMs) with SS-GANS with the goal of mitigating the DOIC problem and experimental results indicating that the architecture achieves the goal.

Abstract

Generative adversarial networks (GANs) have shown tremendous promise in learning to generate data and effective at aiding semi-supervised classification. However, to this point, semi-supervised GAN methods make the assumption that the unlabeled data set contains only samples of the joint distribution of the classes of interest, referred to as inliers. Consequently, when presented with a sample from other distributions, referred to as outliers, GANs perform poorly at determining that it is not qualified to make a decision on the sample. The problem of discriminating outliers from inliers while maintaining classification accuracy is referred to here as the DOIC problem. In this work, we describe an architecture that combines self-organizing maps (SOMs) with SS-GANS with the goal of mitigating the DOIC problem and experimental results indicating that the architecture achieves the goal. Multiple experiments were conducted on hyperspectral image data sets. The SS-GANS performed slightly better than supervised GANS on classification problems with and without the SOM. Incorporating the SOMs into the SS-GANs and the supervised GANS led to substantially mitigation of the DOIC problem when compared to SS-GANS and GANs without the SOMs. Furthermore, the SS-GANS performed much better than GANS on the DOIC problem, even without the SOMs.

Figures (13)

• Figure 1: SOM weights trained with data from the Santa Barbara data described in Section \ref{['sec:santa_barbara']}
• Figure 2: Histograms showing several materials' response to the SOM in Figure \ref{['fig:santabarbara_SOM']}. These histograms show the frequency each node is the closest node to the samples from that class. Note the similarity in response of the map to the classes which are subspecies of the same species.
• Figure 3: 5x5 SOM weights trained with spectra from the tree and mostly grass classes.
• Figure 4: Average Mahalanobis distance between materials from the MUUFL dataset and SOM nodes. The trees and mostly grass materials are the inlier materials.
• Figure 5: Average spectral angle between materials from the MUUFL dataset and SOM nodes. The trees and mostly grass materials are the inlier materials.