Challenges in OpenSet Learning of Neural Networks
- DOI
- 10.2991/978-94-6239-823-8_4How to use a DOI?
- Keywords
- open set learning; OpenMax; extra class insertion
- Abstract
Standard neural network classifiers assume a closed-world setting, where all test-time classes are known during training. In practice, models often face samples from unseen classes, a challenge known as open-set recognition. We investigate and compare two approaches: OpenMax, which fits a Weibull distribution to penultimate-layer activations to estimate class membership, and an extra-class insertion method, which explicitly models an “unknown” class during training. We evaluate both on Tiny ImageNet, CIFAR-10, MNIST, COIL-20, and the Dry Bean dataset under open-set conditions, with selected classes withheld from training. Results show that OpenMax yields only marginal gains over a closed-set baseline, largely due to unreliable Weibull fitting and invalid feature-space assumptions. Extra-class insertion consistently outperforms OpenMax, nearly doubling accuracy on unknown classes in our experiments, making it a more practical foundation for open-set learning.
- Copyright
- © 2026 The Author(s)
- Open Access
- Open Access This chapter is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/), which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.
Cite this article
TY - CONF AU - Péter Bán AU - László Kovács PY - 2026 DA - 2026/11/04 TI - Challenges in OpenSet Learning of Neural Networks BT - Proceedings of the Innovation in Engineering Education 2026 Conference (IEE 2026) PB - Atlantis Press SP - 29 EP - 40 SN - 2667-128X UR - https://doi.org/10.2991/978-94-6239-823-8_4 DO - 10.2991/978-94-6239-823-8_4 ID - Bán2026 ER -