Abstract
In this chapter, we present CORrelation ALignment (CORAL), a simple yet effective method for unsupervised domain adaptation. CORAL minimizes domain shift by aligning the second-order statistics of source and target distributions, without requiring any target labels. In contrast to subspace manifold methods, it aligns the original feature distributions of the source and target domains, rather than the bases of lower-dimensional subspaces. It is also much simpler than other distribution matching methods. CORAL performs remarkably well in extensive evaluations on standard benchmark datasets. We first describe a solution that applies a linear transformation to source features to align them with target features before classifier training. For linear classifiers, we propose to equivalently apply CORAL to the classifier weights, leading to added efficiency when the number of classifiers is small but the number and dimensionality of target examples are very high. The resulting CORAL Linear Discriminant Analysis (CORAL-LDA) outperforms LDA by a large margin on standard domain adaptation benchmarks. Finally, we extend CORAL to learn a nonlinear transformation that aligns correlations of layer activations in deep neural networks (DNNs). The resulting Deep CORAL approach works seamlessly with DNNs and achieves state-of-the-art performance on standard benchmark datasets.
Introduction
In machine learning, the assumption that training (source) and test (target) data share the same distribution often fails in practice due to "domain shift." For instance, a model trained on high-quality studio images may struggle with low-resolution smartphone photos. While supervised adaptation is possible, labeling new data for every environment is prohibitively expensive. Unsupervised Domain Adaptation (UDA) addresses this by leveraging unlabeled target data.
Existing UDA methods frequently rely on complex subspace manifolds or projection techniques that can be computationally intensive or lose information through dimensionality reduction. CORAL (CORrelation ALignment) offers a "frustratingly easy" alternative: it aligns the second-order statistics (covariances) of source and target distributions directly in the original feature space.
Methods
The paper introduces three variations of the CORAL approach:
Linear CORAL
Linear CORAL is the foundational algorithm. It applies a linear transformation to source features through a two-step process:
- Whitening: It removes the correlations of the source domain, making the covariance matrix an identity matrix.
- Re-coloring: It injects the correlations of the target domain into the whitened source features.
This process ensures the transformed source distribution matches the target distribution's feature correlations while preserving the original feature dimensionality.
CORAL-LDA and Deep CORAL
For high-dimensional scenarios like object detection, CORAL-LDA applies the transformation to classifier weights rather than individual data points, significantly improving efficiency.
Deep CORAL extends this concept to Deep Neural Networks (DNNs) by introducing a "CORAL loss" layer. This layer calculates the Frobenius norm of the difference between source and target covariance matrices. By optimizing a joint loss—Classification Loss plus CORAL Loss—the network learns features that are both discriminative for the task and invariant across domains.
Key Results and Insights
CORAL was evaluated on standard benchmarks including Office-Caltech10 and the Standard Office Dataset.
- Performance: On shallow features (Office-Caltech10), Linear CORAL achieved 46.7% accuracy, outperforming TCA and Subspace Alignment. In deep learning, Deep CORAL reached 72.1% accuracy on the Office dataset, surpassing other deep adaptation methods like DAN and DDC.
- Information Preservation: Unlike subspace methods that require choosing a $k$-dimensional basis, CORAL works with original features, avoiding information loss and hyperparameter tuning for $k$.
- Stability: A significant advantage of CORAL is its robustness. As shown in Figure 3, the method maintains stable performance across a wide range of regularization parameters ($\lambda$), making it highly practical for real-world deployment.
Conclusion
CORAL demonstrates that aligning second-order statistics is a powerful and efficient strategy for unsupervised domain adaptation. While it currently focuses on mean and covariance, future work could explore combining CORAL with other alignment losses to capture higher-order distribution structures. Its simplicity, stability, and state-of-the-art performance make it a robust baseline for domain adaptation tasks.
Reference: Sun, B., Feng, J., & Saenko, K. (2016). Correlation Alignment for Unsupervised Domain Adaptation. arXiv:1612.01939v1 [cs.CV].
Summary updated