independent-research

Home Independent Research

Index

Saito, Kuniaki, et al. “Open Set Domain Adaptation by Backpropagation.” arXiv preprint arXiv:1804.10427 (2018).

Motivation

The paper proposed a method for an open set domain adaptation scenario which utilizes adversarial training.

Method

The overview method looks like:

where there are two networks, a feature generator G followed by a classifier C, with two objective functions.

The proposed method firstly trained both classifier and the generator to categorize source samples correctly.

where x_s is the input source and y_s is the output source.

And another classifier is trained to make a boundary for an unknown sample, which is target samples (unlabeled data). The paper proposed to utilize a binary cross entropy loss.

Then the overall objective functions for classifier and feature generator look like:

The classifier attempts to minimize both loss function whereas the generator attempts to maximize the value of L_adv(x_t) to deceive the classifier, such that it can generator better features which would recognize “known” samples from unlabeled target samples.