Topological Signatures Boost Hyperdimensional Computing Robustness

A recent study on arXiv (2605.16785) advances hyperdimensional (HD) computing by identifying discrete topological features—specifically holes—from binarized forms and associating them with shape signatures that are invariant to rotation, translation, and scaling (RTS). While HD computing serves as a lightweight substitute for deep networks in edge learning, traditional pixel-based encoders struggle with distribution shifts such as rotation, noise, or occlusion. The new technique creates RTS-stable descriptors: employing a spatial-pyramid variant of Zernike moments for the outer shape and an intrinsic Fourier descriptor for each hole's radial signature with RTS-canonical geometry. Each feature is converted into a bipolar hypervector through randomized projection and role binding, with variable-cardinality hole sets combined using permutation-invariant bundling. This innovation addresses a significant limitation of HD computing, enhancing its applicability in real-world edge scenarios where robustness is essential.