Build the positive pair
Image embedding
Paired text target
Image embedding → text target
LeVLJEPA objective lab
Explore the idea described in the post: predict a paired text embedding from an image embedding, update on that positive pair, and never construct a negative batch.
1 paired image1 text target0 negativesloss = 1 − cosinepredict, compare, update1 paired image1 text target0 negativesloss = 1 − cosinepredict, compare, update
One pair. One target. A measurable update.
This is an objective explorer, not the LeVLJEPA research model. Its transparent vector math demonstrates the zero-negative mechanism described in the source.
Observed objective
Cosine similarity0.0000
Positive-pair loss0.0000
Negatives used0
Ready for a predictor step.
Predictor vector
Update the prediction
Invariant check
The objective compares one prediction with its paired text target.
- Positive targets1
- Negative targets0
- Contrastive rankingnone
- Predictor steps0
The image representation does not compete against a gallery of wrong captions. It moves toward the one paired text target, making the learning signal visible without negatives.
Predict
Begin with an image-side predictor vector. The lab keeps every coordinate visible so the experiment can be inspected and exported.
Compare
Measure cosine similarity against only the selected paired text target. Loss is computed as one minus that similarity.
Update
Move each predictor coordinate toward its target by the chosen learning rate. No negative example enters the calculation.
Keep the result, not just the animation.
Export the selected pair, raw vectors, update settings, before-and-after measurements, and zero-negatives invariant as a reproducible JSON record.