I would like to use CNN for regression in keras package (backend theano).
I have the training data which has 3-dimension and numbers in each coordinate. And the data also corresponds to certain number.
e.g.)one traing data..
x_train=numpy.arange(0,27).reshape(3,3,3), y_train=0.73
x_train corresponds to y_train
My problem is that sample size is smaller than number of coordinates.
As far as I researched, sparse modeling seems to be useful.
Is it correct? If so, teach me how to do sparse modeling in keras package.
And if you know more information in my problem, please teach me.
Suggest you start with lasso which handles the p>n case. See the attached files on how to do that esp. werk... Section on hi dimension data. This is in R. Google R for Dummies PDF to download a basic R reference. I t tells you how to use and download R. Contact me with any questions. Best regards, David Booth
Hi Ryosaku,
I'd recommend that you try to further pre-process your data before you start modeling. If you have many more coordinates than samples, you could perform principle components analysis to reduce the dimensionality of your dataset (refer to: https://stats.stackexchange.com/questions/28909/pca-when-the-dimensionality-is-greater-than-the-number-of-samples#28914).
In my opinion, the keras documentation is pretty good and you might find an answer to your question if you look at NN models for sparse matrices. Another interesting post in this regard is: https://stackoverflow.com/questions/39977511/training-keras-neural-network-with-sparse-data#40622267.
I am not sure if that answers your question, though. You might want to give a more detailed explanation of how your dataset looks like. I guess if the overall sample size is also small, regularized regression (e.g. after PCA) as recommended above could be a viable alternative to a neural network.
Best,
Daniel
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression