Domain adaptation with subspace learning

Transductive Learning Via Improved Geodesic Sampling

In this paper, we aim to sample more intermediate space between the source and target domain and improve the performance of transductive learning.


  • We propose an improved Geodesic Sampling model (GSM) to generalize prior work to include Riemannian manifolds that can be potentially more useful for transductive learning.

  • We present geodesic sampling formulations and algorithms to apply GSM to sphere and Kendall’s shape manifolds. For highest absolute performance, we also create two updated datasets (Office + Caltech 10, Office 31 and OfficeHome) for domain adaptation whose features are extracted using a pre-trained Xception deep neural network.

  • Our empirical results find that the much simpler sphere manifold often works better than the more commonly adopted Grassmannian manifolds and encourages more efficient computation.

Generate more intermediate subspace (black points on the red curve) with GSM

Extracted features from pre-trained Xception network

t-SNE view of four domains in Office + Caltech 10 dataset

Results

Domain adaptation for object recognition using subspace sampling demons

In this paper, we overcome the limitation of proposed geodesic sampling on manifold (GSM) model, that the points are sampled in an analogous linear manner, but does not generate problematic deformations as shown in


  • We propose a subspace sampling demon (SSD) approach for fast intermediate feature learning.

  • We provide a quantitative evaluation of learned features, so that we are able to select the best features for the image recognition problem. Evaluation of learned features can be of value to future research on domain adaptation.

  • To better represent the learned features and train a robust classifier, we align both marginal and conditional distributions of source and target domains.

Shape Deformation comparison of our proposed Subspace sampling demons (SSD), GSM and previous sampling geodesic sampling (SGF) from Gopalan et al. 2011.

The square shape is continuously deformed in our SSD model, while both GSM and SGF only show a linear changed from a square to a circle, and SGF model does not generate a correct sample in (c)

Results