A high performance suite of SVD related solvers for machine learning

2020-02-05T05:42:33Z (GMT) by Zhenming Liu Andreas Stathopoulos
We present our recent research progress on joint-optimization for ML algorithm and SVD solver. Our major discovery is that the "stopping criteria" of a SVD algorithm can be directly optimized for downstream ML applications. We will present a few examples, in which when we change the stopping criteria of the "inner SVD algorithm", we see significant performance gain (in running time) in the "outer" ML algorithm.