Martingale-Consistent Self-Supervised Learning Framework Introduced
A groundbreaking self-supervised learning (SSL) framework has been introduced by researchers, which ensures martingale consistency, allowing for coherent predictions between coarse and refined views. Unlike traditional SSL objectives that focus on invariance and clustering representations, this innovative method only constrains the expected refined prediction, enabling updates while preserving coherence. The framework features practical variants for both prediction and latent space, along with an unbiased two-sample Monte Carlo estimator derived from stochastic refinement. It has been tested on various benchmarks, including synthetic and real time-series, tabular, and image data. This research is documented in arXiv preprint 2605.11846.
Key facts
- Self-supervised learning often deployed under changing information (shorter histories, missing features, partially observed images).
- Martingales formalize coherence principle for predictions from coarse and refined views.
- Standard SSL objectives do not enforce martingale consistency.
- New framework closes gap with prediction- and latent-space variants.
- Unbiased two-sample Monte Carlo estimator based on stochastic refinement introduced.
- Evaluated on synthetic and real time-series, tabular, and image benchmarks.
- Preprint available on arXiv with ID 2605.11846.
- Approach allows predictions to update as information is revealed while preventing systematic drift.
Entities
Institutions
- arXiv