How are the features obtained in a sparse autoencoder?

https://www.mathworks.com/help/nnet/examples/training-a-deep-neural-network-for-digit-classification.html

In the above tutorial, how do we get the image features in the first hidden layer?

This is a homework question and I can't seem to figure out how exactly the trainAutoencoder function is carrying out the feature extraction. Like, it has to go through some feature detection, followed by a feature extraction algorithm, right? Is that what it's doing?

NOTE: The original question is: How were the features in Fig 3 obtained? Fig 3 refers to the features learned by the autoencoder representing curls and stroke patterns from the digit images.

Answer 1

[BERGHOUT Tarek]

in spearse autoencoders , a set of the original images mapped to the output layer passing by the hidden layer, where the outputs inintialy is the same as the input (g(H)=x) and H is the hidden layer.

but in sparse auto encoder the hidden layer is not the as hidden layer in ordinary autoencoder; the hidden layer must be 'sparse': contains the maximam number of Zeros, that is mean we will code the input with only the significant features in the hidden layer.

thers is an example here : http://ufldl.stanford.edu/wiki/index.php/Exercise:Sparse_Autoencoder

go check it.