MATLAB: How can we change the cost function of Autoencoder

Question

I want to change the cost function of Autoencoder in MATLAB instead of the traditional cost function that minimizes the reconstruction error. Can I do it directly by explicitly mentioning it somewhere? OR do I have to go in the code of Autoencoder. Can you please suggest me the location where I may have to make changes?

Answer 1

I'm guessing you would have to change the code itself, since there doesn't appear to be any way of changing the cost function . Specifically, the 'LossFunction' parameter passed to trainAutoencoder appears to have 'msesparse' as the only option available.

However, you can change 2 coefficients of the cost function: the 'L2WeightRegularization' coefficient and the 'SparsityRegularization' coefficient . The form of the L ₂ Regularization and Sparsity Regularization terms appear fixed, though. It's just their relative weighting that is changed in the cost function.