How to create a layer using Neural Network like the C3 Layer of LeNet?

the C3 Layer of LeNet-5,is not a fully connect convolutional layer. every filter of this layer only take 4~5 feature map as a input from the previous layer. how could I build this kind of layer on Matlab? Using the Neural Network ToolBox or other function?

Answer 1

[Carl]

Hi Lingjun. Using the Neural Network Toolbox, you can create a convolutional neural net, and specify the number of filters to apply at each convolutional layer. The number of filters is analogous to the number of output channels, or feature maps. The following page has a good explanation of how this fits into the overall conv net structure:

https://www.mathworks.com/help/nnet/ug/layers-of-a-convolutional-neural-network.html

And here is the documentation for the convolution2dLayer function that you can use to specify the convolutional layer and the feature maps:

https://www.mathworks.com/help/nnet/ref/convolution2dlayer.html

[LINGJUN KONG]

so, does which mean I cannot custom a cross feature map connection like the S2,C3 Layers in LeNet5?

[Carl]

Thank you for the follow up, I seem to have misunderstood your original question. As of MATLAB R2017a, this feature is not supported. See the documentation here on the 'NumChannels' property:

https://www.mathworks.com/help/nnet/ref/convolution2dlayer.html#input_argument_d0e56199

It states that "this parameter is always equal to the channels of the input to this convolutional layer," meaning that any subsequent layer will be connected to all the channels (feature maps) of the previous layer.

[LINGJUN KONG]

Thanks for your reply. Hope the Matlab Group could let this software more suitable for ML