Tile 2D
This layer groups neighboring pixels into a single super pixel to reduce the size of the image without losing any information. This layer is also known as block layer in Alex Grave's RNNLIB and the seminal NIPS paper "Offline Handwriting Recognition with Multidimensional Recurrent Neural Networks" by A. Graves and J. Schmidhuber.
This layer has a fairly straightforward implementation. The only detail is that shared memory was used in the GPU implementation.
The image below shows how the pixels are grouped into super pixels using a 2x2 tile. Each pixel value is added as a new channel in the output image. When the image is not divisible by the size of the kernel, zero-padding is used to complete the image.
