TS2Vec-FM freezes a 1.6M-parameter time-series encoder
Original title: Turning TS2Vec into a Foundation Model for Time Series Classification
Kept frozen, it scores 83.3% on UCR-128 and 75.3% on UEA-27, within 0.3 points of MantisV2.
A paper argues that TS2Vec, normally retrained per dataset, can become a foundation model with three changes: synthetic-corpus pre-training, dual raw-amplitude and instance-normalized shape stems, and distillation into one encoder. The result, TS2Vec-FM, is a 1.6M-parameter dilated convolutional network without attention or a patch tokenizer. Under linear probing it beats TS2Vec trained separately on each target dataset and sits within 0.3 percentage points of MantisV2.
- TS2Vec was introduced in 2022 as a per-dataset self-supervised method.
- The encoder accepts variable-length inputs without resampling.
- Reported scores are 83.3% on UCR-128 and 75.3% on UEA-27.
- Near-parity uses a convolutional encoder and a two-layer MLP stem.
From the original article
1 min read at arxiv.orgTS2Vec is one of the strongest self-supervised methods for time series classification, introduced in 2022. It is designed to be trained anew on every dataset: self-supervised pre-training followed by linear probing. Since then, the field has shifted towards foundation models that reuse one encoder across datasets.
By Vasilii Feofanov · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.