Reproduce “Deep CORAL: Correlation Alignment for Deep Domain Adaptation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Baochen Sun, Kate Saenko

The paper introduces Deep CORAL, an end-to-end unsupervised domain adaptation method that integrates correlation alignment into deep neural networks. It achieved state-of-the-art average accuracy on the Office dataset by minimizing the Frobenius norm of the covariance difference between source and target features via a differentiable loss function, enabling effective non-linear domain transfer.

Key insights

  • Aligning second-order statistics (covariances) within a deep feature space via a differentiable loss is an effective strategy for mitigating domain shift in an end-to-end deep learning framework.
  • Deep CORAL enables the learning of non-linear transformations for domain adaptation, moving beyond the linear limitations of the original CORAL method.
  • Balancing the classification objective with domain alignment through a weighted loss function allows the network to maintain discriminative power while becoming robust to domain differences.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.