Domain Adaptation

Making a model trained on one data distribution (the source domain) work well on a different distribution (the target domain).

Domain adaptation is a branch of transfer learning: the source domain has plenty of labeled data, while the target domain has a different distribution and little or no labeling, and the goal is to make the model work on the target domain too. Two classic approaches exist: at the feature level, such as Ganin and Lempitsky's 2014 gradient reversal layer, which trains a classifier that tries to tell which domain the data came from and reverses its gradient, forcing the model to learn features that are indistinguishable across domains; and at the pixel level, using a GAN (generative adversarial network) to translate source-domain images into the target domain's visual style. In embodied AI, the most typical case is simulation-to-real: simulation data is cheap but differs from reality, and domain adaptation and domain randomization are the two main ways to narrow that sim-to-real gap.

ExampleGoogle's GraspGAN (2017) used pixel-level domain adaptation to make simulated grasping images look more like real camera images; the paper reports this cut the number of real samples needed to match the same grasping performance by up to 50-fold.

Also called

Unsupervised Domain Adaptation

Related

Transfer Learning · Sim-to-Real Transfer · Sim-to-Real Gap (Reality Gap) · Domain Randomization · Distribution Shift · Out-of-Distribution

Sources

Unsupervised Domain Adaptation by Backpropagation (arXiv:1409.7495)
Using Simulation and Domain Adaptation to Improve Efficiency of Deep Robotic Grasping (arXiv:1709.07857)