TabTalk #1: Introducing MANTIS: a lightweight, calibrated foundation model for user-friendly time series classification.

Mantis is an open-source transformer-based foundation model for time series classification pre-trained exclusively on synthetic data via self-supervised contrastive learning. We argue that effective tokenization is critical, while the full potential of the model can be unlocked at inference time using dedicated strategies. Through extensive experiments we show that Mantis outperforms existing foundation models both when the backbone is frozen and when fine-tuned.Vasilii will present Mantis, the process behind the model's development, and share the advantages of using this model compared to other approaches.

‍

Vasilii Feofanov is a machine learning researcher specializing in foundation models, time series, and transfer learning. He holds a PhD from Université Grenoble Alpes and previously served as Senior Researcher at Huawei Noah's Ark Lab in Paris. His most notable work includes SAMformer and Mantis, both addressing how foundation models can be made practical and lightweight for real-world time series applications.Vasilii has built his career at the intersection of theoretical ML and applied research. He is currently based in Paris and recently joined 42.com as Founding AI Research Scientist.