Code for custom CNNs on custom dataset, relative to the paper "How deep convolutional neural networks lose spatial information with training".

https://scholar.google.com/citations?view_op=view_citation&hl=it&user=cQqn1TQAAAAJ&citation_for_view=cQqn1TQAAAAJ:qjMakFHDy7sC

scale_detection_1d

Training simple CNNs on the Scale-Detection task. Compute their sensitivities to input transformations.

#create_dataset

Creates the Scale-Detection dataset in 1 dimension.

It is possible to choose the characteristic length \xi and the image size, and whether Task 1 and Task 2.

#convolutional_nets

Simple CNNs used to learn Scale-Detection task, with ReLU non linearity.

Parameters: filter size, stride, number of channels, depth.

#ScaleDetection_SGD

Training simple CNNs on the Scale-Detection task.

input parameters:

characteristic scale

image size

gap

number of layers

learning rate

number channels

magnitude active pixels in the dataset

ridge for weight decay

p for the weight decay p-norm

#diffeo_cluster_mean

Computes network stability to diffeomorphisms and Gaussian noise, layer by layer.

Possible to choose whether averaging all the channels within a layer or not.