FDSML: Formula-Driven Supervised Metric Learning with Parameter-Aware Triplet Loss

Researchers proposed Formula-Driven Supervised Metric Learning (FDSML), a pre-training method that replaces the classification-on-discretized-labels approach used in Formula-Driven Supervised Learning (FDSL) with metric learning, preserving continuity in the generative process. The work introduces a parameter-aware triplet loss and derives relations from formula-generated data to enable vision pre-training without real images. The method targets the information loss FDSL incurs by discarding continuity when discretizing labels for classification.

Formula-Driven Supervised Learning FDSL enables vision pre-training without real images, but existing methods cast pre-training as classification on discretized labels, discarding continuity in the generative process. We propose Formula-Driven Supervised Metric Learning FDSML , which derives rela