Domain Separation Networks

By Konstantinos Bousmalis, George Trigeorgis, Nathan Silberman, Dilip Krishnan, Dumitru Erhan. 2016

1,611 citations recorded by the source corpus.

The cost of large scale data collection and annotation often makes the application of machine learning algorithms to new tasks or datasets prohibitively expensive. One approach circumventing this cost is training models on synthetic data where annotations are provided automatically. Despite their appeal, such models often fail to generalize from synthetic to real images, necessitating domain adaptation algorithms to manipulate these models before they can be successfully applied. Existing approaches focus either on mapping representations from one domain to the other, or on learning to extract features that are invariant to the domain from which they were extracted. However, by focusing only on creating a mapping or shared representation between the two domains, they ignore the individual characteristics of each domain. We hypothesize that explicitly modeling what is unique to each domain can improve a model's ability to extract domain-invariant features. Inspired by work on private-shared component analysis, we explicitly learn to extract image representations that are partitioned into two subspaces: one component which is private to each domain and one which is shared across domains. Our model is trained to not only perform the task we care about in the source domain, but also to use the partitioned representation to reconstruct the images from both domains. Our novel architecture results in a model that outperforms the state-of-the-art on a range of unsupervised domain adaptation scenarios and additionally produces visualizations of the private and shared representations enabling interpretation of the domain adaptation process.

Subfields: Artificial Intelligence, 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

In Representation & Self-Supervised Learning

Citations in the corpus

Learning Robust Global Representations by Penalizing Local Predictive Power (2019, methodology): “…adversarial (or similar) techniques (Ghifary et al., 2015; Wang et al., 2016; Motiian et al., 2017b; Li et al., 2018a; Carlucci et al., 2018), or build ensembles of per-domain models that are then fused representations together (Bousmalis et al., 2016; Ding and Fu, 2018; Mancini et al., 2018).”
Adversarial Discriminative Domain Adaptation (2017, background): “Domain separation networks [15] enforce these adversarial losses to minimize domain shift in a shared feature space, but achieve impressive results by augmenting their model with private feature spaces per-domain, an additional dissimilarity loss between the shared and private spaces, and a reconstr”
Continual and Multi-Task Architecture Search (2019, background): “However, strictly imposing this condition into the objective function is not feasible (Bousmalis et al., 2016), hence we add a penalizing term into the objective function as an approximation to the orthogonality condition: Both Condition 2.1 and 2.2 are mutually dependent, because for two matrices’…”
On the Transfer of Inductive Bias from Simulation to the Real World: a New Disentanglement Dataset (2019, background): “4 First Experimental Evaluations of (unsupervised) Disentanglement Methods on Real-World Data Some fields have been able to narrow the gap between simulation and reality [56, 5, 23], which has led to remarkable achievements (e.”
Deep learning predicts hip fracture using confounding patient and healthcare variables (2018, neutral): “62) Fracture frequency, No. (%) 572 (3) 207 (3) 207 (50) 207 (51) 207 (50) 207 (50) BMI, mean (SD) 28 (7) 28 (7) 25 (5) 25 (5) 24 (5) 24 (4) Fall frequency, No. (%) 3,214 (18) 1,139 (19) 133 (32) 160 (40) 174 (42) 165 (40) Pain frequency, No. (%) 9,010 (51) 2,960 (50) 164 (39) 137 (34) 117 (28) 104 ”
Sim-to-Real Robot Learning from Pixels with Progressive Nets (2016, background): “Several recent works from the supervised learning literature, e.g. [4, 1, 3], demonstrate how ideas from the adversarial training of neural networks can be used to reduce the sensitivity of a trained network to inter-domain variations, without requiring aligned training data.”
Adversarial Continual Learning (2020, background): “orthogonality constraints introduced in [30], also known as “difference” loss in the domain adaptation literature [4], to prevent the shared features between all tasks from appearing in the private encoded features.”
Domain Adaptation with Asymmetrically-Relaxed Distribution Alignment (2019, methodology): “Concerning algorithms, several prior works demonstrate empirical success of domain-adversarial approaches, (Tzeng et al., 2014; Ganin et al., 2016; Bousmalis et al., 2016; Tzeng et al.; Hoﬀman et al., 2017; Shu et al., 2018).”
Moment Matching for Multi-Source Domain Adaptation (2018, background): “decoder [3, 10] or a GAN discriminator, such as dualGAN [46], cycle-GAN [50], disco-GAN [16], and CyCADA [15].”
Data Decisions and Theoretical Implications when Adversarially Learning Fair Representations (2017, background): “In particular, researchers have included an adversary to help compensate for skewed data distributions in domain adaptation problems for robotics and simulations [3, 6].”
Unsupervised Pixel-Level Domain Adaptation with Generative Adversarial Networks (2016, methodology): “[5] introduce a model that explicitly separates the components that are private to each domain from those that are common to both domains.”
Efficient Parametrization of Multi-domain Deep Neural Networks (2018, background): “[4] consider an explicit parameterization of domain-generic and domain-specific that learns to extract image representations from the partitioned subspaces.”
Using Simulation and Domain Adaptation to Improve Efficiency of Deep Robotic Grasping (2018, background): “a domain-invariant feature extractor, often represented by a convolutional neural network (CNN) [22], [23], [24].”
CyCADA: Cycle-Consistent Adversarial Domain Adaptation (2017, methodology): “Recent methods can learn to translate images given only unsupervised data from both domains (Bousmalis et al., 2016; 2017; Liu & Tuzel, 2016b; Shrivastava et al., 2017).”
Learning to Generalize: Meta-Learning for Domain Generalization (2017, positive): “It has also been achieved based on gradient reversal domain confusion losses in deep networks (Ganin and Lempitsky 2015; Bousmalis et al. 2016).”
Deeper, Broader and Artier Domain Generalization (2017, positive): “Note that DSN is already shown to outperform the related [9].”
Latent Multi-Task Architecture Learning (2017, methodology): “Bousmalis et al. (2016) use shared and private encoders regularized with orthogonality and similarity constraints for domain adaptation for computer vision.”