This page scores scalar toy embeddings rather than learning an encoder from data. A real shared encoder maps inputs to embeddings through shared parameters; manual point edits here isolate the objective. Only A–P and A–N are scored. The fixed pair labels represent authored similarity relationships; distance does not relabel a pair.

Each coordinate is an integer 0..100. Margin is strictly positive, 10..100 in steps of 10. One-dimensional Euclidean distance is |a−b|. The convention is Y=0 similar, Y=1 dissimilar, with factor ½ in each term. Other implementations may reverse label coding or omit that factor; inspect their declared formula. The reference sums pair losses; this page also explicitly averages its two selected terms.

Collapsed puts all points at 40 with Margin 30: the similar penalty is zero, dissimilar penalty ½×30²=450 and mean 225. The negative term penalizes collapse here. The page draws no gradient arrows and makes no claim about a unique descent direction at a coincident dissimilar pair. Zero loss on two authored pairs does not establish learned semantics or generalization.

This margin objective is distinct from triplet loss and from temperature-based multi-negative softmax objectives. No temperature, normalization, probability model, optimizer or automatic training is used. Coordinates, labels and the chosen pair set define every displayed value. Prediction and Reset restore Near negative for the current experiment; free-exploration Reset begins Experiment 1.