Predefined layers

[accosmin]

Layers with no parameters to learn:

• normalize: transforms input x to (x - avg(x)) / stdev(x)
• normalize image per plane: linear transformation to span the full range [0,1]
• histogram equalization
• LBP/MCT: filter each input map|plane using the LBP or the MCT operator
• MB-LBP/MCT: expands each input map using the LBP or the MCT operator at different scales
• per-plane gradients (vertical + horizontal)
• per-plane downsample: to efficiently process large datasets (e.g. for initial testing)
• random Gabor-like convolutions of given kernel size (e.g. 5x5)
• random convolutions of given kernel size (e.g. 5x5)
• random affine layers (see random kitchen sinks)
• local contrast normalization in a given window size

[accosmin]

In the end