![[2505.11139] Covariance Density Neural Networks](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2505.11139] Covariance Density Neural Networks
Summary
The paper introduces Covariance Density Neural Networks (CDNN), enhancing graph neural networks by using the sample covariance matrix as a Graph Shift Operator, improving performance in applications like Brain Computer Interface (BCI) classification.
Why It Matters
This research addresses the challenge of selecting appropriate graph structures in neural networks, particularly for complex data like EEG signals. By leveraging covariance matrices, CDNNs offer a novel approach to improve stability and performance in real-world applications, making them significant for advancements in BCI technology.
Key Takeaways
- CDNNs utilize the sample covariance matrix as a Graph Shift Operator for improved modeling.
- The approach enhances discriminability and robustness to noise in neural networks.
- CDNNs outperform existing models like EEGnet in BCI applications, particularly in subject-independent scenarios.
- The method allows explicit control over the stability-discriminability trade-off.
- This research contributes to the transferability of BCIs across different individuals.
Computer Science > Machine Learning arXiv:2505.11139 (cs) [Submitted on 16 May 2025 (v1), last revised 22 Feb 2026 (this version, v2)] Title:Covariance Density Neural Networks Authors:Om Roy, Yashar Moshfeghi, Keith Smith View a PDF of the paper titled Covariance Density Neural Networks, by Om Roy and 2 other authors View PDF HTML (experimental) Abstract:Graph neural networks have re-defined how we model and predict on network data but there lacks a consensus on choosing the correct underlying graph structure on which to model signals. CoVariance Neural Networks (VNN) address this issue by using the sample covariance matrix as a Graph Shift Operator (GSO). Here, we improve on the performance of VNNs by constructing a Density Matrix where we consider the sample Covariance matrix as a quasi-Hamiltonian of the system in the space of random variables. Crucially, using this density matrix as the GSO allows components of the data to be extracted at different scales, allowing enhanced discriminability and performance. We show that this approach allows explicit control of the stability-discriminability trade-off of the network, provides enhanced robustness to noise compared to VNNs, and outperforms them in useful real-life applications where the underlying covariance matrix is informative. In particular, we show that our model can achieve strong performance in subject-independent Brain Computer Interface EEG motor imagery classification, outperforming EEGnet while being faster. Th...