Introduction
Electroencephalogram (EEG) signals are notoriously difficult to process because they are non-linear, non-stationary, and exhibit significant variability across subjects and sessions. A powerful approach to represent this data is through Symmetric Positive Definite (SPD) matrices, such as covariance matrices, which reside on curved Riemannian manifolds. While Deep Metric Learning (DML) has been applied to classify these matrices, traditional methods typically attempt to learn a single global distance metric for the entire dataset.
The authors of this paper argue that a one-size-fits-all metric is insufficient for the inherent complexity of brain signals. Such a global approach often fails to capture the subtle, localized relationships between different types of neural activity, leading to sub-optimal classification performance and slow training convergence.
Methods: The "Separate-to-Learn" Strategy
To address these limitations, the researchers introduced Riemannian Embedding Banks (REB). This framework integrates with existing Riemannian networks, such as SPDNet, to implement a "separate-to-learn" strategy.
Key Innovations
- Separate-to-Learn (SL) Layer: Instead of mapping all data into a single high-dimensional embedding, REB partitions the embedding space into $K$ non-overlapping subsets, or "banks." Each bank is assigned a dedicated "learner" (a sub-network or neuron) specialized for a specific cluster of the data.
- Riemannian-Aware Clustering: To avoid the high computational cost of iterative geodesic calculations required by standard K-means on manifolds, REB uses a custom objective function. This allows data to self-organize into informative clusters dynamically during training.
- Distance-Weighted Negative Sampling: To prevent model collapse and ensure robust learning, the authors employ a specific probability distribution for sampling negative pairs based on their Riemannian distance. This ensures the network consistently encounters a diverse range of "hard" and "easy" training examples.
- Concatenated Embeddings: The final feature vector is formed by concatenating the outputs of all $K$ learners, leveraging the specialized knowledge captured by each bank for final classification.
Key Insights and Results
The REB framework was evaluated on three major EEG benchmarks: BCICIV (motor imagery), EEGBCI (large-scale motor imagery), and DEAP (emotion analysis).
Spatial Specialization
Qualitative analysis revealed that the individual learners within the bank naturally evolved to focus on distinct brain regions. For instance, one learner might specialize in signals from the centro-parietal lobes, while another focuses on the frontal cortex. This spatial specialization allows the model to disentangle complex neural patterns more effectively than a global metric.
Performance and Convergence
REB demonstrated significant improvements over standard metric learning approaches:
- Accuracy Boost: The method achieved an average 14.2% improvement in Mean Average Precision (mAP) compared to standard random sampling.
- State-of-the-Art Results: On the BCICIV dataset, REB with Margin Loss reached 51.2% mAP, outperforming other SOTA methods like DSML (45.9%) and DWL (48.8%).
- Faster Convergence: The "separate-to-learn" architecture enables the model to reach high accuracy in significantly fewer training epochs.
Robustness to Data Imbalance
REB proved particularly effective at handling unbalanced datasets. In tests on the DEAP-9 dataset, standard SOTA methods experienced a performance drop of approximately 14.8%, whereas REB's performance only decreased by 7.7%. By dividing the problem into sub-tasks, the model prevents majority classes from completely overwhelming the learning signals of minority classes.
Conclusion
Riemannian Embedding Banks offer a scalable and efficient solution for the challenges of EEG signal classification. By replacing a single global metric with a bank of specialized learners, REB achieves higher accuracy, faster convergence, and greater robustness to non-stationary and unbalanced data. Future work may explore methods to automatically determine the optimal number of embedding banks during the training process.
Summary updated