Energy Efficient Explainable Regularization Technique for Sustainable Trustworthy AI

We introduce Geometric Shallow Learning (GSL) as an alternative to Geometric Deep Learning (GDL) by demonstrating how Hypersphere Neural Networks (HNNs) leverage geometric inductive biases in a shallow yet structured way. While GDL relies on deep architectures to extract geometric features, GSL explicitly encodes these features into a constrained model. This work extends hypersphere-based decision boundaries by introducing a novel regularization technique that improves generalization, energy efficiency, and interpretability. Unlike prior works on hypersphere neurons (HN) that focus on their basic classification capabilities, we introduce a graph-regularized modular approach to training, inspired by GDL concepts like equivariance, steerability, and manifold adaptation. Additionally, we benchmark our model against standard Multi Layer Perceptron(MLP) to validate energy efficiency and interpretability. Our results show that HNNs trained with our new regularization, not only converge faster, but also demonstrate better energy consumption compared to standard MLPs with traditional L1/L2 regularization.