Adaptive Ensemble Embedding for Transfer Learning

In this work, we present an approach for transfer learning. Many adversarial transfer learning approaches train merely one domain classifier to differentiate the features as either source domain or target domain and train a feature generator to perplex the discriminator. One problem exists with this family of approaches, where the single discriminator is used to differentiate the features as source or target thus could not mine the hard samples. To solve the above problem, we introduce a new adaptive ensemble embedding scheme for transfer learning (AEETL). We propose to learn multiple adaptive embedding features with considering the hard samples for transfer learning. AEETL mainly contains two components: (1) the weight module is applied to predict the adaptive ensemble weights; (2) the ensemble module is applied to mine the hardness of samples following a coarse-to-fine scheme. Comprehensive experimental results demonstrate the effectiveness and superiority of our approach.