Search International and National Patent Collections

1. (WO2017148314) METHOD OF TRAINING MACHINE LEARNING SYSTEM, AND TRAINING SYSTEM

Pub. No.:    WO/2017/148314    International Application No.:    PCT/CN2017/074318
Publication Date: Sat Sep 09 01:59:59 CEST 2017 International Filing Date: Wed Feb 22 00:59:59 CET 2017
IPC: G06F 17/30
G06N 5/00
Applicants: ALIBABA GROUP HOLDING LIMITED
阿里巴巴集团控股有限公司
ZHOU, Jun
周俊
Inventors: ZHOU, Jun
周俊
Title: METHOD OF TRAINING MACHINE LEARNING SYSTEM, AND TRAINING SYSTEM
Abstract:
A method of training a machine learning system, and a training system, utilizing training data to train the machine learning system. The method comprises: assigning, to a plurality of workstations, training data (S101); dividing the training data assigned to each of the workstations into a plurality of data slices (S102); acquiring a local weight and a local loss function value computed by each of the workstations on the basis of each of the data slices (S103); collecting the local weight and the local loss function value to obtain a current weight and a current loss function value (S104); utilizing the current weight and/or the current loss function value to perform a model abnormality check (S105); when a check result indicates a first type of abnormality, inputting into the machine learning system the previously collected weight and the previously collected loss function value to perform training (S106); and when the check result indicates a second type of abnormality, modifying the current weight and/or the current loss function value to be within a first threshold, and inputting the same into the machine learning system (S107).