Mantis: towards a powerful foundation model for time series classification

Foundation models for time series data have recently gained attention for their potential to generalize across diverse tasks. However, most existing models focus on forecasting, leaving time series classification underexplored despite its importance in areas like human activity recognition and medical diagnostics. We introduce
Mantis, an open-source, transformer-based foundation model pre-trained with contrastive learning for time series classification. Lightweight and resource-efficient, Mantis can serve as a feature extractor or be fine-tuned for downstream tasks. Empirical results show that Mantis outperforms existing time series foundation models, offering an
effective and versatile solution.

Bio: Vasilii Feofanov, Ph.D., is a senior research scientist at Huawei Noah’s Ark Lab. His recent research focuses on time series classification, the efficiency of transformers for time series, and out-of-distribution (OOD) generalization of pre-trained models. He completed his Ph.D. at Grenoble Alpes University, where he worked on
the theoretical and practical aspects of semi-supervised learning with a particular focus on self-training algorithms.