Paper characterizes JEPA feature selection as predictive closure
Original title: Predictive Closure: A Set-Level Characterization of JEPA's Feature Selection
Unpredictable features can still be encoded when they help predict other features.
The paper reformulates the JEPA training objective so feature selection depends on sets of features rather than each feature's individual predictability. This predictive-closure view agrees with earlier accounts when features evolve independently, but predicts that a hard-to-predict feature such as a traffic light can be retained if it helps predict other features. Controlled experiments support that claim and show steering one feature can strengthen representation of another.
- Selection depends on properties of feature sets, not only individual predictability.
- Changing one feature's predictive links changed how others were represented.
- An auxiliary objective or side information can steer related features indirectly.
- Experiments were controlled rather than on a production JEPA system.
From the original article
1 min read at huggingface.co
In joint-embedding predictive architectures (JEPA), an encoder and a predictor are trained together to produce a representation that minimizes the error in predicting the representation of the next observation. Because the representations of both the current and the next observation are learned, it is unclear which features they will include.
By Mosh Levy, Aly Lidayan, Yoav Goldberg · huggingface.co
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.