Analysis of (sub-)Riemannian PDE-G-CNNs (Q6186783)

The article proposes a group equivariant convolutional neural network based on partial differential equations (PDE-based G-CNNs) which, compared to the canonical convolutional neural networks, reduce network complexity, enhance classification performance and provide geometric interpretability. The limitations of traditional morphological kernels are outlined and new approximative kernels that address issues arising from spatial anisotropy are proposed. The authors validate their approach with theorems (accompanied by proofs), but also through empirical tests on two datasets. PDE-G-CNNs reduces the complexity of the networks on the two problems, while maintaining, or even improving, performance compared to traditional CNNs and G-CNNs, and having the advantage of the geometric interpretability. The findings are certainly promising and, if the wider research community can validate its effectiveness across different real-world tasks, the approach may well set a new standard for deep learning models.