hongyusu / L2RTA

http://hongyusu.github.io/L2RTA This repository contains RTA algorithm with improved inference algorithm for update direction finding. The RTA algorithm was originally introduced in Advances in Neural Informatics Processing System NIPS 2014. In each stochastic gradient descent iteration, the new inference algorithm aims to combine multiple updated directions. In other words, the optimal updated direction is given by a convex combination of many suboptimal update directions. Thus, the new inference algorithm avoid computing the best direction from the K-best list which was used in original RTA algorithm.