TS2Vec-FM
A time series foundation model improved from TS2Vec that adopts a dilated convolution architecture with only 1.6M parameters and supports linear probing with a frozen encoder, approaching state-of-the-art on several time series classification benchmarks.
Timeline (last 90 days)
TS2Vec-FM is a dilated convolutional network with only 1.6M parameters.
Under a frozen-encoder, linear-probe-only setting, TS2Vec-FM achieves 83.3% accuracy on UCR-128.
Under the frozen-encoder linear probing setting, TS2Vec-FM surpasses TS2Vec trained separately on each target dataset.
TS2Vec-FM turns TS2Vec into a foundation model through three key modifications: pretraining on a synthetic corpus, a dual-view input design (raw amplitude view and instance-normalized shape view), and fusion into a single encoder via distillation.
TS2Vec-FM does not use attention mechanisms, does not use a patch tokenizer, and does not rely on intermediate-layer feature extraction.
TS2Vec-FM achieves near-optimal performance with only a convolutional encoder and a two-layer MLP stem.
All Facts (6)
TS2Vec-FM is a dilated convolutional network with only 1.6M parameters.
50%UnverifiedUnder a frozen-encoder, linear-probe-only setting, TS2Vec-FM achieves 83.3% accuracy on UCR-128.
50%UnverifiedUnder the frozen-encoder linear probing setting, TS2Vec-FM surpasses TS2Vec trained separately on each target dataset.
50%UnverifiedTS2Vec-FM turns TS2Vec into a foundation model through three key modifications: pretraining on a synthetic corpus, a dual-view input design (raw amplitude view and instance-normalized shape view), and fusion into a single encoder via distillation.
50%UnverifiedTS2Vec-FM does not use attention mechanisms, does not use a patch tokenizer, and does not rely on intermediate-layer feature extraction.
50%UnverifiedTS2Vec-FM achieves near-optimal performance with only a convolutional encoder and a two-layer MLP stem.
50%