wangchaohahh/tensorflow-ddt

tensorflow-DDT

Most prior domain adaptation methods only focus on reducing the difference in the marginal distribution between domains. An emerging question, however, is that the MMD value is quite small, but the two distributions are indistinguishable due to the small variance and inter-class distance. We instead use optimized MMD to ensure that the full feature distributions discrepancy is reduced, and discriminative property for classification tasks is reserved. This helps avoid the common failure of domain adaptation models, where the whole domain data collapses to outputting the indistinguishable situation of different categories.

experiment result

Figure.2 shows that the target categories are discriminated better by the target classifier (larger class-to-class distances), which suggests that our method is a reasonable extension to previous deep feature adaptation methods.