Adversarial latent representation for positive unlabeled learning
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: generative adversarial networks, representation learning, novelty detection, transductive learning, positive unlabeled learning
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Novelty detection, a widely studied problem in machine learning, is the task of detecting a novel class of data that has not been previously observed. Deep networks have driven the state-of-the-art work on this application in recent years due to their successful applications on large and more complex datasets. The usual setting for novelty detection is unsupervised whereby only examples of the normal class are available during training, but more recently there has been a surge in interest in semi-supervised methods. A common assumption about semi-supervised methods is their access to an additional set of labeled data that includes a few examples of anomalies. Transductive novelty detection or positive-unlabeled (PU) learning on the other hand assumes access to an additional unlabeled set that contains examples of anomalies. In this study, we focus on machine vision applications and propose TransductGAN, a transductive generative adversarial network (GAN) that attempts to learn how to generate image examples from the novel class by separating the latter from the negative class in a latent space using a mixture of two Gaussians. It achieves that by incorporating an adversarial autoencoder with a GAN network; the ability to generate examples of novel data points offers not only a visual representation of the new class, but also overcomes the hurdle faced by many inductive methods about how to tune the model hyperparameters at the decision rule level. In addition, the introduction of a latent space enables an enhanced discriminative learning. Our model has shown superior performance over state-of-the-art work on several benchmark datasets.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5582
Loading