horn-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Edward J. Yoon (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HORN-21) Implement dropout for training
Date Mon, 20 Jun 2016 11:56:05 GMT

    [ https://issues.apache.org/jira/browse/HORN-21?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15339385#comment-15339385
] 

Edward J. Yoon commented on HORN-21:
------------------------------------

I implemented dropout neuron and tested with input layer drop rate 0.8 and hidden layer 0.5,
and high learning rate and momentum. I was able to reach 0.95 accuracy within only few minutes
using 20 tasks on my mac pro. Training time will be reduced to 1/3 on 64 cores server.

I will commit my changes if there's no objection within 3 days.

> Implement dropout for training
> ------------------------------
>
>                 Key: HORN-21
>                 URL: https://issues.apache.org/jira/browse/HORN-21
>             Project: Apache Horn
>          Issue Type: New Feature
>          Components: api
>            Reporter: Minho Kim
>
> http://arxiv.org/pdf/1207.0580.pdf
> Dropout is a relatively algorithm which appears to be highly effective for improving
the quality of neural network. It is a vital feature in almost neural network implementation.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Mime
View raw message