Averaging Stochastic Gradient Descent on Riemannian Manifolds

Citations Over Time

Abstract

We consider the minimization of a function defined on a Riemannian manifold $\mathcal{M}$ accessible only through unbiased estimates of its gradients. We develop a geometric framework to transform a sequence of slowly converging iterates generated from stochastic gradient descent (SGD) on $\mathcal{M}$ to an averaged iterate sequence with a robust and fast $O(1/n)$ convergence rate. We then present an application of our framework to geodesically-strongly-convex (and possibly Euclidean non-convex) problems. Finally, we demonstrate how these ideas apply to the case of streaming $k$-PCA, where we show how to accelerate the slow rate of the randomized power method (without requiring knowledge of the eigengap) into a robust algorithm achieving the optimal rate of convergence.

Related Papers

• → Prim�rzerlegung in Steinschen Algebren(1964)32 cited
• → A historical overview of pro cdh descent in algebraic $K$-theory and its\n relation to rigid analytic varieties(2016)1 cited
• → Descent and Landing Performance(2012)
• Susquehanna Chorale Spring Concert "Roots and Wings"(2017)
• → ИСПОЛЬЗОВAНИЕ ПОТЕНЦИAЛA СОЦИAЛЬНЫХ ПAРТНЕРОВ В ПОДГОТОВКЕ БУДУЩИХ ПЕДAГОГОВ(2024)