I know about the structure of a convolutional neural net and I have even applied it to the MNIST digit recognition problem.
I was looking at a cloud based deep learning API and it asks about specifying the number of piecewise linear maxout units in my convolutional neural network?
As far as I understand, a typical convnet has a series of convolution kernels , each followed by a sub-sampler with an optional pooling layer, before the next convolutional layer starts. Then, at the end we have a fully connected hidden layer of (may be sigmoid neurons) followed by the output layer (which could be softmax).
Where in the above structure does a piecewise linear maxout unit fit? I know what a piece wise linear maxout unit does mathematically, but where does it fit in the above structure? Does it replace any of the above components or is it an addition at a particular phase?