[
https://issues.apache.org/jira/browse/SYSTEMML-2018?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16259672#comment-16259672
]
Mike Dusenberry commented on SYSTEMML-2018:
-------------------------------------------
[~return_01] Yeah that would be interesting to explore further. I would
suggest creating a new optimizer called {{AdamW}} as is done in the paper,
rather than altering the current Adam one.
> Fixing Weight Decay Regularization in ADAM
> ------------------------------------------
>
> Key: SYSTEMML-2018
> URL: https://issues.apache.org/jira/browse/SYSTEMML-2018
> Project: SystemML
> Issue Type: Improvement
> Components: Algorithms
> Reporter: Janardhan
>
> The common implementations of adaptive gradient algorithms, such
> as Adam, limit the potential benefit of weight decay regularization, because
> the
> weights do not decay multiplicatively (as would be expected for standard
> weight
> decay) but by an additive constant factor.
> This following paper found a way to fix regularization in Adam Optimization
> with one addition step(+ wx) to the gradient step :
> https://arxiv.org/pdf/1711.05101.pdf
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
