Turning TS2Vec into a Foundation Model for Time Series Classification
arXiv cs.AI · 21 Oct 2026, 03:47 AEST
TS2Vec has been adapted into a foundation model for time series classification, retaining its compact architecture. The new model, TS2Vec-FM, achieves competitive performance while using a traditional convolutional encoder.
Key facts
- TS2Vec is one of the strongest self-supervised methods for time series classification, introduced in 2022.
- Three changes suffice: pre-train on a synthetic corpus, feed the encoder a raw amplitude view and an instance-normalized shape view of the input through two small stems, and distill one specialist encoder per view into a single encoder.
- The result, TS2Vec-FM, is a 1.6M-parameter dilated convolutional network that, kept frozen, reaches 83.3% on UCR-128 and 75.3% on UEA-27 under linear probing.
- What we find most interesting is how different TS2Vec-FM is from other foundation models in design: near-parity is reached without attention, without a patch tokenizer, and without intermediate-layer extraction.
Read aloud
TS2Vec has been adapted into a foundation model for time series classification, retaining its compact architecture. The new model, TS2Vec-FM, achieves competitive performance while using a traditional convolutional encoder. TS2Vec is one of the strongest self-supervised methods for time series classification, introduced in twenty twenty two. Three changes suffice: pre-train on a synthetic corpus, feed the encoder a raw amplitude view and an instance-normalized shape view of the input through two small stems, and distill one specialist encoder per view into a single encoder. The result, TS2Vec-FM, is a one.6M-parameter dilated convolutional network that, kept frozen, reaches eighty three point three percent on UCR-one hundred twenty eight and seventy five point three percent on UEA-twenty seven under linear probing. What we find most interesting is how different TS2Vec-FM is from other foundation models in design: near-parity is reached without attention, without a patch tokenizer, and without intermediate-layer extraction.
Sources in this cluster
- Turning TS2Vec into a Foundation Model for Time Series Classification
arXiv cs.AI · 21 Oct 2026, 03:47 AEST