Semantic Regularization for Source-Free Semi-Supervised Domain Adaptation
SERL adapts a source-pretrained model to a new target domain using only a few target labels and no source data, by adding three complementary regularizers. Semantic probability contrastive regularization compares samples probabilistically with adaptive weights to learn discriminative features; hard-sample mixup regularization blends easy and hard target samples to extract knowledge from hard cases; and target prediction regularization keeps current predictions correlated with previously learned targets to reduce the harm of incorrect pseudo-labels. Together they yield state-of-the-art semi-supervised domain-adaptation accuracy.
D
Dr. Harry Seldon
Text
Source-free Semantic Regularization Learning for Semi-supervised Domain Adaptation Xinyang Huang
SERL (Semantic Regularization Learning) is a framework for semi-supervised domain adaptation that fine-tunes a source-pretrained model on a target domain using only a few target labels and no access …