Introduction
Dictionary learning (DL) is a fundamental technique in machine learning, designed to identify a set of "atoms" or basis functions that can sparsely represent complex data. While well-established for univariate signals, applying DL to multivariate data—such as EEG recordings from dozens of sensors—presents significant challenges. Specifically, comparing learned dictionaries to ground truths or other models has historically relied on heuristic "detection rates." These heuristics are often sensitive to arbitrary thresholds and fail to satisfy the mathematical properties of a true metric, such as the triangle inequality. This paper proposes a rigorous framework for dictionary comparison by treating multivariate atoms as points on a Grassmannian manifold and applying Wasserstein-like set-metrics.
Methodology
The authors bridge the gap between Frame Theory and Differential Geometry through a three-step approach:
- Subspaces over Matrices: Rather than treating multivariate atoms as raw matrices, they are modeled as subspaces. This ensures that comparisons are invariant to linear transformations, such as rotations, which do not alter the underlying information content.
- Grassmannian Manifolds: The space of all such subspaces is defined as a Grassmannian manifold ($Gr(\rho, n)$). Ground metrics, such as chordal or geodesic distances, are used to measure the distance between individual subspaces on this manifold.
- Wasserstein Set-Metrics: To compare entire dictionaries (sets of atoms), the authors lift these ground metrics using Wasserstein distances (Earth Mover's Distance). This provides a global, smooth average distance that is more robust than outlier-sensitive measures like the Hausdorff distance.
Experimental Validation
The proposed metrics were validated using both synthetic and real-world datasets:
- Synthetic Datasets: Dictionaries with $m=135$ atoms and $ ho=10$ components were used to test recovery under various conditions. The Wasserstein-chordal metric demonstrated smooth, monotonic convergence over iterations, whereas traditional detection rates appeared "jumpy" and sensitive to noise.
- BCI Competition IV-2a (EEG): This benchmark dataset contains EEG recordings from subjects performing motor imagery tasks. The metrics were used to quantify intersession variability, providing mathematical evidence for the "drift" in brain signals that necessitates daily recalibration of Brain-Computer Interface (BCI) systems.
Key Insights and Neurobiological Alignment
A significant finding of this work is the link between geometry and neurobiology. When clustering dictionaries from different motor imagery tasks, the metrics naturally grouped "hand" and "tongue" imagery together while separating "feet" imagery. This alignment mirrors the human homunculus, where the cortical representations of the hand and tongue are physically closer to each other than to the feet.
Furthermore, the metrics identified "BCI-inefficient" subjects as geometric outliers. This suggests that these subspace metrics could serve as diagnostic tools for assessing dataset quality or subject-specific signal characteristics.
Conclusion and Future Work
By providing a mathematically sound framework for comparing multivariate dictionaries, this research offers a more reliable way to track algorithm convergence and analyze complex signals like EEG. While the computational complexity of Wasserstein distances is higher than simple heuristics, the benefits of smoothness and geometric invariance are substantial. Future research may involve integrating these metrics directly into the loss functions of dictionary learning algorithms to enforce geometric diversity or physiological priors during the learning process.
Summary updated