Efficient Learning of Domain-invariant Image Representations

By Judy Hoffman, Erik Rodner, Jeff Donahue, Kate Saenko, Trevor Darrell. 2013

296 citations recorded by the source corpus.

Abstract: We present an algorithm that learns representations which explicitly compensate for domain mismatch and which can be efficiently realized as linear classifiers. Specifically, we form a linear transformation that maps features from the target (test) domain to the source (training) domain as part of training the classifier. We optimize both the transformation and classifier parameters jointly, and introduce an efficient cost function based on misclassification loss. Our method combines several features previously unavailable in a single algorithm: multi-class adaptation through representation learning, ability to map across heterogeneous feature spaces, and scalability to large datasets. We present experiments on several image datasets that demonstrate improved accuracy and computational advantages compared to previous approaches.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, 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 Computer Vision and Pattern Recognition

In Representation & Self-Supervised Learning

Citations in the corpus

Domain randomization for transferring deep neural networks from simulation to the real world (2017, background): “The computer vision community has devoted signiﬁcant study to the problem of adapting vision-based models trained in a source domain to a previously unseen target domain (see, e.g., [9], [14], [15], [19], [23], [25], [51]).”
Neural Discrete Representation Learning (2017, background): “At the same time, challenging tasks such as few-shot learning [34], domain adaptation [17], or reinforcement learning [35] heavily rely on learnt representations from raw data, but the usefulness of generic representations trained in an unsupervised fashion is still far from being the dominant…”
One-Shot Imitation Learning (2017, background): “Success stories include domain adaptation in computer vision [64, 34, 28, 4, 15, 24, 33, 59, 14] and control [60, 45, 46, 20, 54].”
Agnostic Federated Learning (2019, background): “…unsupervised adaptation techniques (Gong et al., 2012; Long et al., 2015; Ganin and Lempitsky, 2015; Tzeng et al., 2015), or via lightly supervised ones (some amount of labeled data from the target domain) (Saenko et al., 2010; Yang et al., 2007; Hoffman et al., 2013; Girshick et al., 2014).”
Algorithms and Theory for Multiple-Source Adaptation (2018, background): “…et al., 2012, Long et al., 2015, Ganin and Lempitsky, 2015, Tzeng et al., 2015], or via lightly supervised ones (some amount of labeled data from the target domain) [Saenko et al., 2010, Yang et al., 2007, Hoffman et al., 2013, Girshick et al., 2014], has been extensively investigated in the past.”
Cross Modal Distillation for Supervision Transfer (2015, background): “Initial work [26, 14, 1, 8, 20] studied the problem in context of shallow image representations.”
Simultaneous Deep Transfer Across Domains and Tasks (2015, methodology): “In Table 2 we present multi-class accuracies over the 16 held-out categories and compare our method to a previous domain adaptation method [18] as well as a source-only trained CNN.”
Deep Domain Confusion: Maximizing for Domain Invariance (2014, background): “Recently, supervised convolutional neural network (CNN) based feature representations have been shown to be extremely effective for a variety of visual recognition tasks [24, 11, 15, 30].”
Third-Person Imitation Learning (2017, background): “…of adapting what is learned in one domain to another domain has been studied extensively in computer vision in the supervised learning setting Yang et al. (2007); Mansour et al. (2009); Kulis et al. (2011); Aytar & Zisserman (2011); Duan et al. (2012); Hoffman et al. (2013); Long & Wang (2015).”
LSDA: Large Scale Detection through Adaptation (2014, background): “, feature space transformations [12], model adaptation approaches [13, 14] and joint feature and model adaptation [15, 16].”
Deep Hashing Network for Unsupervised Domain Adaptation (2017, methodology): “One straightforward approach is, to modify a classifier trained for the source data by adapting it to classify target data [1, 4] or learn a transformation matrix to linearly transform the source data, so that it is aligned with the target [27, 42].”
One-Shot Adaptation of Supervised Deep Convolutional Models (2013, methodology): “In particular, MMDT minimizes the following loss function (assuming a binary classification task to simplify notation, and with `hinge defined as in PMT): LMMDT (θ, A) = 1 2 ‖θ‖22 + 1 2 ‖A− I‖2F + Cs`hinge(X,Y ;θ) + Ct`hinge(AX̃, Ỹ ;θ) , (2) where Cs and Ct are hyperparameters controlling the impor”
Domain Adaptation for Visual Applications: A Comprehensive Survey (2017, methodology): “The PMT-SVM was further combined by Donhahue et al. in [64] with the Max-Margin Domain Transforms [117] to handle complex domain shifts.”
Transfer from Simulation to Real World through Learning Deep Inverse Dynamics Model (2016, methodology): “Our approach is agnostic to how the policy πsource was obtained.”
Adapting Deep Visuomotor Representations with Weak Pairwise Constraints (2015, methodology): “Later models improved upon this by proposing adaptation which both transformed the representation spaces and regularized the target model using the source data [19,20].”
PixelVAE++: Improved PixelVAE with Discrete Prior (2019, background): “Many machine learning tasks such as domain adaptation [12, 13, 14, 15] and few-shot learning [16, 17, 18] rely on learned representations of observations.”
Deep Classifiers from Image Tags in the Wild (2015, unknown): “, [9]) could also be used.”