LogHD: Robust Compression of Hyperdimensional Classifiers via Logarithmic Class-Axis Reduction