Joint cross-domain classification and subspace learning for unsupervised adaptation

By Basura Fernando, Tatiana Tommasi, Tinne Tuytelaars. 2014

60 citations recorded by the source corpus.

Focus on unsupervised domain adaptation by subspace learning.Highlight the importance of learning jointly the prediction model and the subspace.Propose the JCSL method.Thorough experimental evaluation against existing subspace adaptive methods.Insight into JCSL with analysis on parameters and domain shift reduction. Domain adaptation aims at adapting the knowledge acquired on a source domain to a new different but related target domain. Several approaches have been proposed for classification tasks in the unsupervised scenario, where no labeled target data are available. Most of the attention has been dedicated to searching a new domain-invariant representation, leaving the definition of the prediction function to a second stage. Here we propose to learn both jointly. Specifically we learn the source subspace that best matches the target subspace while at the same time minimizing a regularized misclassification loss. We provide an alternating optimization technique based on stochastic sub-gradient descent to solve the learning problem and we demonstrate its performance on several domain adaptation tasks.

Subfields: Artificial Intelligence, Cancer Research, Epidemiology, Representation & Self-Supervised Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Saeid Motiian
2. Gianfranco Doretto
3. Barbara Caputo
4. Q. Jones
5. Seyed Mehdi Iranmanesh
6. Wouter M. Kouw
7. Marco Loog
8. Tatiana Tommasi

In Cancer Research

1. Gianfranco Doretto
2. Saeid Motiian
3. Marco Loog
4. Oncel Tuzel
5. Wouter M. Kouw
6. Ming-Yu Liu
7. Mathieu Salzmann
8. Pascal Fua

In Epidemiology

1. Gianfranco Doretto
2. Saeid Motiian
3. Marco Loog
4. Oncel Tuzel
5. Wouter M. Kouw
6. Ming-Yu Liu
7. Pascal Fua
8. Artem Rozantsev

In Representation & Self-Supervised Learning

1. Barbara Caputo
2. Wouter M. Kouw
3. Gianfranco Doretto
4. Saeid Motiian
5. Marco Loog
6. Oncel Tuzel
7. Ming-Yu Liu
8. Mathieu Salzmann

Citations in the corpus

• Coupled Generative Adversarial Networks (2016, methodology): “Method [17] [18] [19] [20] CoGAN From MNIST 0.”
• Few-Shot Adversarial Domain Adaptation (2017, methodology): “The MNIST (M), USPS (U), and SVHN (S) datasets have recently been used for domain adaptation [12, 45, 59].”
• Unified Deep Supervised Domain Adaptation and Generalization (2017, methodology): “We considered two cross-domain tasks,M→ U and U →M, and followed the experimental setting in [20, 50], which involves randomly selecting 2000 images from MNIST and 1800 images from USPS.”
• A Review of Domain Adaptation without Target Labels (2019, methodology): “Transfer Subspace Learning minimizes the Bregman divergence to both domains, with respect to a projection to a subspace [145].”
• Learning the Roots of Visual Domain Shift (2016, methodology): “Few of them rely on SVM-like risk minimization objectives and shallow representation models [12,13], while more recent approaches leverage over deep learning [14,15,3].”
• Multi-Domain Adversarial Learning (2019, methodology): “Other approaches proceed by aligning the source and target representations with PCA-based correlation alignment (Sun et al., 2016), or subspace alignment (Fernando et al., 2015).”
• Beyond Sharing Weights for Deep Domain Adaptation (2016, methodology): “They have recently been employed for the task of Domain Adaptation [68].”
• Information Bottleneck Domain Adaptation with Privileged Information for Visual Recognition (2016, background): “those we mention [39, 16] because they do so based on information theoretic measures, like we do.”