Rahul Iyer created MADLIB-1272:
----------------------------------
Summary: Simplify Nesterov by performing velocity update in
previous iteration
Key: MADLIB-1272
URL: https://issues.apache.org/jira/browse/MADLIB-1272
Project: Apache MADlib
Issue Type: Bug
Components: Module: Neural Networks
Reporter: Rahul Iyer
Momentum updates are complicated due to Nesterov requiring an initial update
before gradient calculations. There is, however, a different form of the
Nesterov update that can be cleanly performed after the regular update,
simplifying the code. This allows performing the gradient calculations before
any update - with or without Nesterov.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
