[mlpack-svn] [MLPACK] #345: Sparse Autoencoder Module
MLPACK Trac
trac at coffeetalk-1.cc.gatech.edu
Tue Apr 15 11:31:59 EDT 2014
#345: Sparse Autoencoder Module
----------------------------+-----------------------------------------------
Reporter: siddharth.950 | Owner:
Type: enhancement | Status: new
Priority: major | Milestone:
Component: mlpack | Resolution:
Keywords: | Blocking:
Blocked By: |
----------------------------+-----------------------------------------------
Comment (by rcurtin):
Yeah -- there is a sparsity penalty parameter, but unless I have
misunderstood it does not actually force the activations to be 0, just
small. What I mean is that you end up with activations that are like
1e-10, but not 0. The model, as you have it implemented, will represent
them as 1e-10; then, when the model is used, all of the connections
between the hidden layer and the visible layer are evaluated. If it was
truly sparse, many of those connections could be ignored, and execution
could probably be faster. That's an idea for another day, though.
However... it does beg the question (and I'm not sure how I overlooked
this before): you've implemented support for training a sparse
autoencoder, but what will a user be able to do with this object once it's
trained? I don't see any functions that would allow the user to actually
utilize the result to make predictions or anything like that.
--
Ticket URL: <http://trac.research.cc.gatech.edu/fastlab/ticket/345#comment:10>
MLPACK <www.fast-lab.org>
MLPACK is an intuitive, fast, and scalable C++ machine learning library developed at Georgia Tech.
More information about the mlpack-svn
mailing list