Deep Optimal Transport for Domain Adaptation on SPD Manifolds

Ce Ju, Cuntai Guan

TL;DR

Deep Optimal Transport (DOT) is a new geometric deep learning framework designed to solve the "session-to-session" problem in brain-computer interfaces (BCIs). By treating EEG covariance matrices as points on a curved Symmetric Positive Definite (SPD) manifold, DOT uses optimal transport theory to align data distributions across different recording sessions, significantly improving classification accuracy while respecting the data's intrinsic geometry.

What Problem Does This Paper Address?

In the world of neuroimaging, particularly for Brain-Computer Interfaces (BCIs), researchers face a major hurdle: domain shift. Brain signals (EEG) are notoriously non-stationary; a model trained on a user’s data from Monday might perform poorly on Tuesday because the user’s mental state, electrode placement, or environmental noise has changed.

Technically, these EEG signals are often represented as spatial covariance matrices. These matrices don't live in standard "flat" Euclidean space; they reside on Symmetric Positive Definite (SPD) manifolds. Conventional domain adaptation methods often ignore this curved geometry, treating the matrices like ordinary vectors. This "flat" approach leads to suboptimal performance because it violates the mathematical properties (symmetry and positive definiteness) inherent to the data.

Methods

The authors propose Deep Optimal Transport (DOT), which integrates the SPDNet architecture with Optimal Transport (OT) theory.

Key components include:

  • SPDNet Encoder: Uses specialized layers like BiMap (linear transformation), ReEig (non-linear rectification), and LogEig (mapping to a flat tangent space) to process covariance matrices while staying on the manifold.
  • Log-Euclidean Metric (LEM): The framework adopts the LEM to measure distances. Unlike other Riemannian metrics, LEM provides a "flat" vector space structure in the logarithmic domain, making it computationally efficient for deep learning gradients.
  • Composite OT Loss: Instead of just minimizing classification error, DOT minimizes a joint loss function:
  • Cross-Entropy Loss: For classification accuracy.
  • Marginal Distribution Adaptation (MDA): Aligns the overall "cloud" of source and target data.
  • Conditional Distribution Adaptation (CDA): Aligns data class-by-class (e.g., ensuring "left-hand" imagery in the source matches "left-hand" imagery in the target) using pseudo-labels for the unlabeled target data.
The DOT pipeline: from multi-channel signals to manifold-aware distribution alignment.

Datasets

The authors validated DOT on three prominent cross-session EEG motor imagery datasets:

  1. KU (Korea University) Dataset: 54 subjects, binary-class task.
  2. BNCI2014001 (BCIC-IV-2a): 9 subjects, four-class task (left hand, right hand, feet, tongue).
  3. BNCI2015001: 12 subjects, binary-class task.

Key Insights and Ideas

  • Geometry Matters: The paper demonstrates that respecting the SPD manifold is not just a mathematical nicety—it's a performance necessity. Methods using the Log-Euclidean Metric consistently outperformed those using standard Euclidean approximations.
  • LEM vs. AIRM: While the Affine-Invariant Riemannian Metric (AIRM) is theoretically popular, the authors argue it is unsuitable for deep learning because it requires iterative optimization that breaks the backpropagation chain. LEM is the "Goldilocks" metric: geometrically sound yet computationally friendly.
  • Joint Alignment: Aligning only the marginal distribution (the whole data cloud) isn't enough. By adding Conditional Distribution Adaptation (CDA), the model ensures that the internal class structures are preserved during the transfer, which is vital for BCI reliability.
Table 1: A comparison of various OT-DA Frameworks. XS and XT are feature spaces of the source and target domains, respectively. P and Q represent the marginal probability distribution and conditional probability distribution, respectively.

Key Results

DOT consistently outperformed baseline approaches (like SVM, MDRM, and standard SPDNet) across all three datasets.

  • On the BNCI2014001 dataset, traditional Optimal Transport (EMD+SVM) achieved only 50.89% accuracy under one metric, while the proposed approach reached significantly higher levels (up to 72.11% when combined with recentering techniques).
  • The framework showed particular strength in unsupervised domain adaptation, where no labels are available for the new session, making it highly practical for real-world BCI applications.

Limitations and Future Work

The authors identify two main limitations:

  1. Pseudo-label Sensitivity: The CDA loss relies on pseudo-labels for the target domain. If the initial model is weak, these labels are noisy, which can lead to "error propagation" where the model learns from its own mistakes.
  2. Quantifying Shift: There is currently no robust way to measure the intensity of a domain shift before attempting to fix it.

Future work will likely focus on uncertainty-aware alignment to make pseudo-labeling more robust and developing better metrics to quantify how much a user's brain signals have actually shifted between sessions.

Summary updated

Abstract

Recent progress in geometric deep learning has drawn increasing attention from the machine learning community toward domain adaptation on symmetric positive definite (SPD) manifolds, especially for neuroimaging data that often suffer from distribution shifts across sessions. These data, typically represented as covariance matrices of brain signals, inherently lie on SPD manifolds due to their symmetry and positive definiteness. However, conventional domain adaptation methods often overlook this geometric structure when applied directly to covariance matrices, which can result in suboptimal performance. To address this issue, we introduce a new geometric deep learning framework that combines optimal transport theory with the geometry of SPD manifolds. Our approach aligns data distributions while respecting the manifold structure, effectively reducing both marginal and conditional discrepancies. We validate our method on three cross-session brain computer interface datasets, KU, BNCI2014001, and BNCI2015001, where it consistently outperforms baseline approaches while maintaining the intrinsic geometry of the data. We also provide quantitative results and visualizations to better illustrate the behavior of the learned embeddings.
BibTeX
@article{ju2025deep,
  author = {Ce Ju and Cuntai Guan},
  title = {{Deep Optimal Transport for Domain Adaptation on SPD Manifolds}},
  journal = {Artificial Intelligence},
  year = {2025},
  doi = {10.1016/j.artint.2025.104347},
}
Figure