I’m experimenting with a Siamese network using triplet loss to categorize sub-classes into broader classes. My setup differs from traditional triplet loss models: It involves using the sub-class as the anchor and the broader class as the positive (where the sub-class fits) and a different class as the negative (where it doesn’t fit). The goal is to position each sub-class embedding closer to its relevant class and farther from unrelated classes. Would this architecture make sense for capturing context-dependent relationships between sub-classes and classes? Are there any limitations I should be aware of?
I havent written any code. I'm curious about the theorhetical possibility.
