DNN-based Domain Adaptation Methods Sample Clauses

DNN-based Domain Adaptation Methods. Motivated by their reported stunning performances in recent years, we compared performance of GTDA with the performances of a number of Deep Neural Networks-based domain adaptation methods reported on the Office 31 dataset based on ResNet50 features. Specifically, we make comparison with Deep Domain Confusion (DDC) [13] where an objective function including an additional domain confusion term is proposed for learn- ing domain-invariant representations for classification, Deep Adaptation Network (XXX) [14] where more transferable features are learned by adapting source and target distributions in multiple task-specific layers, Residual Transfer Networks (RTN) [11] that achieves feature adaptation and classifier adaptation simultaneously by deep residual learning [28], Reverse Gradient (RevGrad) [12], [15] that improves domain adaptation by employing adversarial training paradigm, and Joint Adaptation Networks (JAN) [16] that uses an adversarial learning strategy to maximize a joint maximum mean discrep- ancy such that distributions of source and target domains be more distinguishable. Despite of high performance accuracies, some disadvantages of DNN-based methods to be taken into account are that they require abundant training data for im- provement in performance, they use target labels for parameter tuning [11] and their sensitivity to hyperparameters of the learning procedure is high [12]. We refer to the performance results reported in [16] for the aforementioned DNN-based techniques to make comparison with our technique, hence we will add results for the Office31 dataset only.