Mantis: Lightweight Calibrated Foundation Model for User-Friendly Time Series Classification
🤖 AI Summary
Addressing the critical shortage of foundational models for time-series classification, this paper introduces Mantis—the first lightweight, calibration-oriented foundation model specifically designed for this task. Built upon the Vision Transformer architecture, Mantis employs contrastive learning for self-supervised pretraining and incorporates a novel multivariate adapter that explicitly models inter-channel dependencies while substantially reducing memory overhead. Its core innovation is a classification-aware prediction calibration mechanism, ensuring optimal calibration performance. Under both frozen-feature and fine-tuning paradigms, Mantis consistently outperforms existing general-purpose time-series foundation models, achieving state-of-the-art accuracy across multiple standard classification benchmarks. Crucially, it maintains efficient inference throughput and low-resource deployment capability. The code and pretrained models are publicly released.
Application Category
📝 Abstract
In recent years, there has been increasing interest in developing foundation models for time series data that can generalize across diverse downstream tasks. While numerous forecasting-oriented foundation models have been introduced, there is a notable scarcity of models tailored for time series classification. To address this gap, we present Mantis, a new open-source foundation model for time series classification based on the Vision Transformer (ViT) architecture that has been pre-trained using a contrastive learning approach. Our experimental results show that Mantis outperforms existing foundation models both when the backbone is frozen and when fine-tuned, while achieving the lowest calibration error. In addition, we propose several adapters to handle the multivariate setting, reducing memory requirements and modeling channel interdependence.
Problem
Research questions and friction points this paper is trying to address.
Develop lightweight model for time series classification
Address scarcity of classification-specific foundation models
Enhance performance with low calibration error
Innovation
Methods, ideas, or system contributions that make the work stand out.
Vision Transformer architecture
Contrastive learning approach
Adapters for multivariate setting
🔎 Similar Papers
No similar papers found.
