Self-Supervised Learning for Time Series: Generative vs Latent Paradigms
A new study from arXiv (2605.19462) systematically compares Generative and Latent Alignment self-supervised learning (SSL) methods for time series, introducing adaptations of LeJEPA and DINO that use Discrete Wavelet Transform (DWT) augmentations. The research finds that the pre-training dividend is highly asymmetric: SSL improves anomaly detection and classification by up to 375%, but offers marginal gains for forecasting. The study establishes a controlled framework to evaluate SSL's value across diverse temporal tasks, challenging the assumption that representational utility is universal.
Key facts
- arXiv paper 2605.19462 compares Generative vs Latent Alignment SSL for time series
- Adapts LeJEPA and DINO using Discrete Wavelet Transform (DWT) augmentations
- SSL yields up to 375% improvement for anomaly detection and classification
- SSL provides marginal improvement for forecasting tasks
- Pre-training dividend is highly asymmetric across tasks
- Controlled framework established to evaluate SSL value for time series
- Research motivated by SSL success in vision and NLP
- Representational utility is non-universal, governed by precision-invariance trade-off
Entities
Institutions
- arXiv