MODEL DISENTANGLEMENT FOR DOMAIN ADAPTATION: abstract simplified (17655506)

The abstract describes a technique for improving domain adaptation in machine learning. It involves generating a feature tensor from input data using a feature extractor. This tensor is then processed by both a domain-agnostic classifier and a domain-specific classifier to generate two sets of logits (outputs). A loss function is computed based on these logits, including a divergence loss component. The feature extractor, domain-agnostic classifier, and domain-specific classifier are then refined using this loss function.