New optimization method achieves curvature-independent regret bounds on manifolds

Researchers have developed a new method for decentralized online optimization on Hadamard manifolds, specifically addressing challenges related to curvature dependence in previous approaches. This work introduces Distributed Riemannian Online Gradient Descent (D-ROGD), which achieves curvature-independent regret bounds for h-convex and strongly h-convex objectives. The method combines local updates with a Fréchet-mean consensus mechanism, yielding static regret rates of O(sqrt(T)) and O(log T) respectively, which match Euclidean rates and are solely dependent on the network's spectral gap. Experimental results on hyperbolic embeddings validate these theoretical findings. AI

IMPACT Introduces theoretical advancements in optimization techniques applicable to machine learning models operating in non-Euclidean spaces.

RANK_REASON Academic paper detailing a new theoretical method for optimization. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• Distributed Riemannian Online Gradient Descent
• Fréchet mean

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →