Multi-Spheres Anomaly Detection with Hyperspherical Layers

This paper investigates the application of hyper-spherical layers in neural networks for anomaly detection, emphasizing Support Vector Data Description (SVDD) and Deep SVDD techniques. We introduce an adaptation of Deep SVDD incorporating a hyperspherical layer defined within conformal geometric algebra. Furthermore, we propose a novel method called Deep M sph-SVDD, which extends this approach to multi-spheres, enabling the model to capture distinct groups of normal data points. We also present new loss functions designed to prevent the intersection and inclusion of spheres. Preliminary experiments on a synthetic dataset are conducted, along with evaluations on the MNIST and CIFAR-IO datasets.