The following is a list of important hyperparameters and their details. Here, I will try to construct a five-layered CNN, as follows:
- Layer 0 has a ConvolutionLayer having a 6 x 6 kernel, one channel (since they are grayscale images), a stride of 2 x 2, and 20 feature maps where ReLU is the activation function:
ConvolutionLayer layer_0 = new ConvolutionLayer.Builder(6,6) .nIn(nChannels) .stride(2,2) // default stride(2,2) .nOut(20) // # of feature maps .dropOut(0.7) // dropout to reduce overfitting .activation(Activation.RELU) // Activation: rectified linear units .build();
- Layer 1 has SubsamplingLayer max pooling, and a stride of 2x2. Thus, by using stride, we down sample by a factor of 2. Note that only MAX, AVG, ...
