[mlpack] pull request for CNN

[Ryan Curtin]

On Thu, Feb 19, 2015 at 11:28:36AM +0800, Shangtong Zhang wrote:
> Hi,
> 
> 
> I want to contribute toEssential Deep Learning Modules in GSoC 2015,
> I want to start with CNN.
> 
> 
> I implement some necessary components and a framewrok for CNN.
> The framework for CNN is very similiar to FFNN, but I make some changes.
> The network in FFNN is  connections, conns, conns, ….., conns 
> I extend it to  layers, conns, layers, conns ….. , layers 
> because I find original structure may lead to duplicate activations in layers in some complicate CNN.
> And my CNN support 2-dim data, so a grey image can be used for input directly.
> 
> 
> I also write a test to prove my CNN does work.
> I implement LeNet1 on part of MNIST dataset(1000 training cases 100 testing cases)
> and achieve 97% recoganization rate on the 100 testing cases.
> LeNet1 is first referred in this paperhttp://yann.lecun.com/exdb/publis/pdf/lecun-90c.pdf
> LeNet1 is a little complicate so I use some feature in c++14 when writing tests and
> executing this test many comsume about 5 minutes.
> 
> 
> I’m looking forward to any feedback and guide.

Thanks!

The C++14 may present a little difficulty -- I am not sure how many
standard compilers support it -- but I'll look into it and see how we
can work around (or work with) it.

I'll comment on the pull request as time permits.

-- 
Ryan Curtin    | "Are those... live rounds?"
ryan at ratml.org | "Seven-six-two millimeter.  Full metal jacket."