Timezone: »

Exponentially convergent stochastic k-PCA without variance reduction
Cheng Tang

Thu Dec 12 10:05 AM -- 10:20 AM (PST) @ West Ballroom A + B

We present Matrix Krasulina, an algorithm for online k-PCA, by gen- eralizing the classic Krasulina’s method (Krasulina, 1969) from vector to matrix case. We show, both theoretically and empirically, that the algorithm naturally adapts to data low-rankness and converges exponentially fast to the ground-truth principal subspace. Notably, our result suggests that despite various recent efforts to accelerate the convergence of stochastic-gradient based methods by adding a O(n)-time variance reduction step, for the k- PCA problem, a truly online SGD variant suffices to achieve exponential convergence on intrinsically low-rank data.

Author Information

Cheng Tang (Amazon)

Related Events (a corresponding poster, oral, or spotlight)