Saito, K., Yamamoto, S., Ushiku, Y., & Harada, T. (2018). Open set domain adaptation by backpropagation. In Proceedings of the European Conference on Computer Vision (ECCV) (pp. 153-168).
Introduction
The proposed method propose a method for an open set domain adaptation scenario, which utilizes adversarial training. This approach allows to extract features that separate unknown target from known target samples.
提出一个检测方式,在不知道source domian的unkown sample的情况下,训练一个网络可以分类出target domain的unkown sample。
在实现过程中有两个问题:
- we do not have knowledge about which samples are the unknown samples. Thus, it seems difficult to delineate a boundary between known and unknown classes.
- need to align target samples with source samples to reduce this domain’s difference, unknown target samples cannot be aligned due to the absence of unknown samples in the source domain
Approach
Unlike the existing distribution alignment methods that only match the source and target distribution, our method facilitates the rejection of unknown target samples with high accuracy as well as the alignment of known target samples with known source samples.
As such, we assign two options to the feature generator: aligning them with samples in the source domain or rejecting them as unknown.
Train a feature generation network , which takes inputs or , and a network , which takes features from and classifies them into classes, where the denotes the number of known categories. The paper propose to train the classifier to output , where 0 < t < 1 and train the generator to deceive the classifier.
根据我的理解,这里有一个分类器和生成器,为了能分类出没有label的unkown objective,作者训练生成器去欺骗分类器,增大error ,分类器更容易去判断一个对象为 unkown class。对于已知类,采用常规的cross entropy如下:
为了能分类出unkown class,利用一个二分类corss entropy:
这里. 总共的训练策略如下:
对于generator而言优化已知类的cross entropy增大unkown class的error,对于classifier而言,同时优化减小已知类的cross entropy和unkown class的error。因此整个优化过程对于unkown class而言是个对抗优化的过程。