Developing a seven learned layer Convolutional Neural Network for Deep Learning

Hello:

I am attempting to replicate a Deep Learned Convolutional Neural Network with seven learned layers for further exploration. The first six layers are the convolutional layers; the final layer is the fully connected layer. The Convolutional layers will include a filters unit, a rectified unit, a pooling unit, and a local normalization unit.

The input of the first layer is a 274 X 274 with 50 filters of size 19 X 19. The second layer will take the outputs from the first layer and filters it with 75 filters of size 11 X 11. The third layer will have a 100 filters of 8 X 8, the fourth layer 250 filters with 5 X 5, the fifth layer with 500 filters of 4 X 4, sixth layer with 2000 filters of 4 X 4; and the seventh layer with 30 filters of 1 X 1.

I plan on training with a Stochastic gradient descent with window size of 25 images; utilizing a data-augmentation technique as well as placing drop-layers with factors of 0.5 on the last two layers to reduce effect over-fitting effects. I plan on the training the CNN with MatConvNet toolbox.

While I have a little experience with developing and training two layer neural networks in PYTHON, I have never attempted a Deep Layer CNN with this many layers in MATLAB (I did not even know Deep Learning could be implemented in this environment) So I am hoping ask the following:

1) Would I have to write up an entire new CNN or could I just modify one of the pre-trained models such as alexNet or googlenet.

2) Is there any sample MATLAB syntax of CNNs with multiple layers for reference? I looked about MathWorks to little success.

3) I was hoping to improve classification transfer learning techniques. Besides mathworks are there any good literature to learn the deeper concepts behind that.

Thank you. I apologize for the long entry. I just wished to over-specific rather than too superficial.

Edit: Almost forgot - Here is my attempt so far

Edit: I don't know why the code won't show properly. Sorry.

 if true

% Convolutional Neural Network for pollen recognition %

layers = [ ... %Convolutional Layers should contain a filter unit, rectified units %(ReLU), pooling unit, and local normalization unit

    %zero-mean Gaussian distribution SHOULD be used to initialize the
    %weights in each layer%

    %a data augmentation technique SHOULD be used for artificially
    %increasing dataset%

    %Stochastic gradient descent SHOULD be used for the training process with
    %window size of 25 images.%

    convolution2dLayer(19, 50) %Convolutional - 1st layer with 50 filters of size 19 X 19
    convolution2dLayer(11, 75) %Convolutional - 2nd layer with 75 filters of size 11 X 11
    convolution2dLayer(8, 100) %Convolutional - 3rd layer with 100 filters of size 8 X 8
    convolution2dLayer(5, 250) %Convolutional - 4th layer with 250 filters of size 5 X 5
    convolution2dLayer(4, 500) %Convolutional - 5th layer with 500 filters of size 4 X 4

    %Drop-out layers by a 0.5 factor SHOULD be attached to last two layers %

    convolution2dLayer(4, 2000)%Convolutional - 6th layer with 2000 filters of size 4 X 4
    fullyConnectedLayer(25)    %Fully Connected - 7th layer SHOULD have 30 filters of size 1 X 1%
    ]
  end