Authors:
András Halász
;
Lóránt Daubner
;
Nawar Al-Hemeary
;
János Juhász
;
Tamás Zsedrovits
and
Kálmán Tornai
Affiliation:
Faculty of Information Technology and Bionics, Pázmány Péter Catholic University, 1083 Práter u. 50/A, Budapest, Hungary
Keyword(s):
Open-Set Recognition, Time-Series.
Abstract:
In real-world scenarios, conventional classifier methods often stumble when faced with the unexpected emergence of unknown samples or classes previously unseen during training. Open-Set Recognition (OSR) models have emerged as a solution to this ubiquitous challenge. Our previous work introduced a robust OSR method leveraging synthesized – or “fake” – features to delineate the uncharted territory of unknowns, focusing on image datasets. Recognizing the imperative to extend this capability to diverse data types, we have successfully transposed this model to time-series datasets. A pivotal feature of the original model was its modular archi-tecture, allowing for focused modification in feature extraction. Consequently, the core components remained intact, including feature extraction, sample generation, and feature transformation. This paper illuminates our initial strides, employing a one-dimensional convolutional network for feature extraction and showcasing promising preliminary OSR
results using that network. Additionally, our adapted model maintains its advantageous edge in terms of time complexity, achieved through the discreet generation of fake features in a simplified hidden layer. Future investigations will further delve into alternative feature extraction methodologies, promising to broaden the scope of applications for this adaptable OSR model.
(More)