In this study we address the critical problem of instability in hyperbolic deep learning, with special attention to learning the curvature of the manifold. Naive techniques present a fundamental weakness identified by the authors: when the curvature parameter is updated before the model parameters, the projections and Riemannian gradients are invalidated and performance deteriorates. To solve this, we propose an ordered projection scheme that preserves the geometric coherence of training. The procedure consists of first updating the curvature, re-projecting the model parameters onto the new manifold, and then projecting them onto a stable tangent space before applying the model parameter update. This order ensures that Riemannian gradients and retractions are valid and improves numerical stability.
In practice, we recommend several complementary strategies to make training more robust: limiting the learning rate of the curvature, using appropriate retractions and parallel transport between manifolds, controlling the numerical condition of the parameterization matrices, and applying geometric regularization that penalizes abrupt curvature jumps. In experiments with hierarchical data and knowledge graphs, the ordered projection scheme shows more stable convergence, better hierarchy representations, and greater tolerance to noise compared to methods that update curvature independently.
The implications for real-world applications are relevant: more stable hyperbolic embeddings improve tasks such as semantic search, graph recommendation, and taxonomy analysis. At Q2BSTUDIO we combine this type of research advances with practical development solutions. If you need to develop solutions around representation models and advanced learning, we offer custom applications and custom software services that integrate geometric models and efficient architectures. We also offer artificial intelligence projects, AI for businesses, and AI agents that incorporate good geometric stability practices to obtain reliable models in production.
We are also specialists in cybersecurity, pentesting, AWS and Azure cloud services, business intelligence services, and Power BI to improve analytics and decision-making. If your project requires integrating hyperbolic models with cloud pipelines, automation, business intelligence services, or security guarantees, at Q2BSTUDIO we design the custom solution that optimizes performance, scalability, and compliance.
Keywords: custom applications, custom software, artificial intelligence, cybersecurity, AWS and Azure cloud services, business intelligence services, AI for businesses, AI agents, Power BI.

.jpg)
