Please enable JavaScript.
Coggle requires JavaScript to display documents.
Metric learning - Coggle Diagram
Metric learning
Types of metric learning
Pairwise ML
Metric learning on a graph
Simple graph: there is an edge between similar pairs
Signed graph: there is a positive (+1) edge between similar pairs and a negative (-1) edge between dissimilar pairs
See also graph layout problem
Metric learning on classes, on cliques
This is a degenerate case of metric learning on a graph where there are only cliques in the graph, corresponding to labels.
A lot of metric learning tasks are framed in this way because it can be done with typical supervised data for classification (i.e. data points (xi, yi) where yi is the class)
-
-
-
-
-
Triplet metric learning
Triplet with class assumption: (positive, anchor, negative)
y(anchor)=y(positive)!=y(negative)
In that case triplet is only a way to leverage the clique information, and there are multiple variation (tuplet loss, subcenter losses...)
Triplet without classes: (A, B, C)
In that case we don't necessarily want y(1) to be close to y(2).
Ex: political opinions, anarchism might be closer to communism than to fascism but are still far away
:bulb: Use geometric intuition: the angle (ABC) must be >90 degrees
-
-
-