| Resumen |
This study aims to address the challenge of generating an appropriate Universum dataset for Universum learning, a learning paradigm that incorporates additional data to a learning problem, with the aim of enhancing a model’s performance on unseen data. We propose a novel approach that generates Universum samples directly from a Variational Autoencoder’s (VAE) learned latent space. In the proposed approach, a VAE is trained with all the available data, labeled and unlabeled. After this, samples are generated using the trained VAE. The generated samples identified as out-of-distribution are used as Universum samples and, along with a few labeled data, are then used to train a Convolutional Neural Network (CNN). The performance of this model is compared with the performance of a vanilla CNN in a multiclass classification problem. Empirical evidence suggests that the samples generated and identified by the VAE as out-of-distribution hold promise as effective tools for generating Universum samples, thereby enhancing model performance and leading to more robust models. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024. |
