Unsupervised Meta-Learning through Latent-Space Interpolation in Generative Models

Unsupervised meta-learning approaches rely on synthetic meta-tasks that are\ncreated using techniques such as random selection, clustering and/or\naugmentation. Unfortunately, clustering and augmentation are domain-dependent,\nand thus they require either manual tweaking or expensive learning. In this\nwork, we describe an approach that generates meta-tasks using generative\nmodels. A critical component is a novel approach of sampling from the latent\nspace that generates objects grouped into synthetic classes forming the\ntraining and validation data of a meta-task. We find that the proposed\napproach, LAtent Space Interpolation Unsupervised Meta-learning (LASIUM),\noutperforms or is competitive with current unsupervised learning baselines on\nfew-shot classification tasks on the most widely used benchmark datasets. In\naddition, the approach promises to be applicable without manual tweaking over a\nwider range of domains than previous approaches.\n

Paper

Similar papers

© 2026 NYSGPT2525 LLC