Convolutional Network

[benoitfragit]

Extend libBrain library with all Convolutin Network object based on this

http://neuralnetworksanddeeplearning.com/chap6.html

http://ufldl.stanford.edu/tutorial/supervised/ConvolutionalNeuralNetwork/

• add an enumeration to save the layer type Convolutional Layer Max Pooling Layer Classical
• Add Max Pooling neuron that take the maximum over a region and attribute the error to the source of error
• See how its possible to ise integral image to reduce array access
• Add several feature to learn by using convolution
• Convolution could be faster with fft:

http://www.silicium628.fr/cours/tr_fourier/page2.php

• Add convolutional layer that apply a patch convolution to a input 2D region. For the backpropagation :

http://andrew.gibiansky.com/blog/machine-learning/convolutional-neural-networks/

Some details about convolutional neural network extracted from this:

• All Convolutional Neuron share the same weights and bias so that they ddetect the same feature
• Max-pooling layer backpropagation attribute all the error to the input that leads to the maximum

[benoitfragit]

C is definitly not the good language to deals with