How can we transfer learnings from training to valid and final?
44 views
Skip to first unread message
JJ
Mar 30, 2011, 7:27:37 PM3/30/11
to Transfer Learning Challenge
Phase 1 and phase 2 have same valid and final dataset, which means if
you can do well in phase 1, you can do well in phase 2 too without
utilizing the transfer learning. You can just fine tune your phase 1
algorithm. But for the sake of transfer learning, how can the
transferred labels be helpful here?
you can do well in phase 1, you can do well in phase 2 too without
utilizing the transfer learning. You can just fine tune your phase 1
algorithm. But for the sake of transfer learning, how can the
transferred labels be helpful here?
Causality Workbench
Mar 30, 2011, 8:36:56 PM3/30/11
to Transfer Learning Challenge
Dear Jianjun,
Did you watch the video posted on the synopsis page:
http://www.youtube.com/watch?v=9ChVn3xVNDI
Was it helpful?
The organizers
Did you watch the video posted on the synopsis page:
http://www.youtube.com/watch?v=9ChVn3xVNDI
Was it helpful?
The organizers
JJ
Mar 31, 2011, 2:42:55 AM3/31/11
to Transfer Learning Challenge
I did watch the video. It is helpful conceptually, not practically. We
tried PCA and clustering in phase 1. Clustering can generate pretty
good learning curve on valid sets. We did k-means directly on valid
set and final set, even not used the devel set. There are two factors
which are critical to clustering: number of clusters, distance matrix.
We got these two factors from the feedback of valid set and used the
same setting for final. The risk is that if the final set has
different number of classes from valid, our approach may fail badly.
does each column represent same variable in devel, vaid and final?
does devel set contain examplars of valid and final?
I can not train a model using labels from devel and apply it to valid
and final since they contain examples for disjoint sets of classes .
If I stay with the clustering approach, I have not seen an easy way to
utilize the transferred labels. I post my puzzles here and hope they
can stimulate more discussions.
> > transferred labels be helpful here?- Hide quoted text -
>
> - Show quoted text -
tried PCA and clustering in phase 1. Clustering can generate pretty
good learning curve on valid sets. We did k-means directly on valid
set and final set, even not used the devel set. There are two factors
which are critical to clustering: number of clusters, distance matrix.
We got these two factors from the feedback of valid set and used the
same setting for final. The risk is that if the final set has
different number of classes from valid, our approach may fail badly.
does each column represent same variable in devel, vaid and final?
does devel set contain examplars of valid and final?
I can not train a model using labels from devel and apply it to valid
and final since they contain examples for disjoint sets of classes .
If I stay with the clustering approach, I have not seen an easy way to
utilize the transferred labels. I post my puzzles here and hope they
can stimulate more discussions.
>
> - Show quoted text -
Causality Workbench
Mar 31, 2011, 12:36:29 PM3/31/11
to Transfer Learning Challenge
One simple way to use the transfer labels is to use them to select
your preprocessing, ie do only unsupervised learning and use the
labels for "model selection" instead or in addition to the performance
on the validation set.
your preprocessing, ie do only unsupervised learning and use the
labels for "model selection" instead or in addition to the performance
on the validation set.
Causality Workbench
Mar 31, 2011, 12:38:13 PM3/31/11
to Transfer Learning Challenge
On Mar 30, 11:42 pm, JJ <jianjun...@gmail.com> wrote:
> I did watch the video. It is helpful conceptually, not practically. We
> tried PCA and clustering in phase 1. Clustering can generate pretty
> good learning curve on valid sets. We did k-means directly on valid
> set and final set, even not used the devel set. There are two factors
> which are critical to clustering: number of clusters, distance matrix.
> We got these two factors from the feedback of valid set and used the
> same setting for final. The risk is that if the final set has
> different number of classes from valid, our approach may fail badly.
>
> does each column represent same variable in devel, vaid and final?
> does devel set contain examplars of valid and final?
Transfer Learning Challenge
Apr 10, 2011, 8:35:38 AM4/10/11
to Transfer Learning Challenge
Hi,
In the ULE dataset, the examples in the devel set are more or less
equally distributed among the different 10 classes, and we can
somewhat infer from the number of transfer labels the number of
classes (26000 examples, 10000 transfer labels for 4 classes, gives
26000/(10000/4)=about 10 classes).
Is this true for the other datasets?
In the ULE dataset, the examples in the devel set are more or less
equally distributed among the different 10 classes, and we can
somewhat infer from the number of transfer labels the number of
classes (26000 examples, 10000 transfer labels for 4 classes, gives
26000/(10000/4)=about 10 classes).
Is this true for the other datasets?
JJ
Apr 12, 2011, 2:07:14 AM4/12/11
to Transfer Learning Challenge
Wow, interesting speculation. Maybe it make sense. Organizer should
know. Not sure she will answer it though.
On Apr 10, 5:35 am, Transfer Learning Challenge <transfer-learning-
> > > > - Show quoted text -- Hide quoted text -
know. Not sure she will answer it though.
On Apr 10, 5:35 am, Transfer Learning Challenge <transfer-learning-
Reply all
Reply to author
Forward
0 new messages