 |
General: 663 |
P2: Escaping from Saddle Points Using Asynchronous Coordinate Descent
Author: Marco Bornstein , Advisor: Furong Huang (CS Department)
Problem Statement Presentation
Abstract
The goal of this project is to implement a first-order asynchronous-coordinate-descent algorithm that has been constructed in parallel, and analyze its convergence. This algorithm is entitled the Saddle Escaping Asynchronous Coordinate Descent (SEACD) algorithm. SEACD is implemented in a parallel manner using Python, and it is compared to another common algorithm: serial gradient descent (GD). The aim is to show that parallel SEACD successfully minimizes non-convex functions, and does so in a quick and efficient manner.
Final Presentation ,
Final Report
