EC-GAN: Low-Sample Classification using Semi-Supervised Algorithms and GANs

Abstract

Semi-supervised learning has been gaining attention as it allows for performing image analysis tasks such as classification with limited labeled data. Some popular algorithms using Generative Adversarial Networks (GANs) for semisupervised classification share a single architecture for classification and discrimination. However, this may require a model to converge to a separate data distribution for each task, which may reduce overall performance. While progress in semi-supervised learning has been made, less addressed are small-scale, fully-supervised tasks where even unlabeled data is unavailable and unattainable. We therefore, propose a novel GAN model namely External Classifier GAN (EC-GAN), that utilizes GANs and semi-supervised algorithms to improve classification in fully-supervised regimes. Our method leverages a GAN to generate artificial data used to supplement supervised classification. More specifically, we attach an external classifier, hence the name EC-GAN, to the GAN's generator, as opposed to sharing an architecture with the discriminator. Our experiments demonstrate that EC-GAN's performance is comparable to the shared architecture method, far superior to the standard data augmentation and regularization-based approach, and effective on a small, realistic dataset. 1

HyperViTGAN: Semisupervised Generative Adversarial Network With Transformer for Hyperspectral Image Classification

“…GANs that share a single architecture for classification and discrimination fail to solve the HSI class imbalance problem properly, and the discriminator is prone to associating fakes with minorityclass(es), which results in weak abilities to classify minorityclass(es) samples. [52] used GAN and a semi-supervised algorithm to supplement the supervised classifier with artificial data. This algorithm, named EC-GAN, was proved to be effective on small, realistic datasets.…”

“…Due to the presence of the weight λ, the fake samples do not contribute much to the model update and classifier loss calculation. We set λ to 0.1 followed [52]. A smaller λ ensures that the model still learns mainly from real HSI samples while the model is fine-tuned using the high-quality fake HSI samples.…”

“…The selection of high-quality fake samples is also critical to the training of the model. We followed the design of [52] and [53] and employed a confidencebased pseudo-labeling scheme. The weak initial generation capability of GAN leads to generation of low-quality samples by the generator.…”

“…As the generative power of GAN gradually improves with training, more and more high-quality fake samples will be adopted into the GAN training process, bypassing the confidence threshold t. The unsupervised loss is obtained by calculating the fake samples and their corresponding pseudo-labels. Here, t is set to 0.7, following [52].…”

RareGAN: Generating Samples for Rare Classes

“…Semi-supervised GANs. Our proposed approach in § 4.1 is one instance of semi-supervised GANs (Odena 2016;Salimans et al 2016;Dai et al 2017;Kumar, Sattigeri, and Fletcher 2017;Haque 2020;Zhou et al 2018). Other semisupervised GANs could also be used, like the seminal one (Salimans et al 2016), which uses a single modified discriminator both to separate fake from real samples (as in classical GANs) and to classify the labels of real data.…”