I want to implement an autoencoder with theano to express each article. When I made it following the tutorial of UFLDL. I found that it's hard to get gradient descent(I use LBFGS). I want to know whether it is due to the sparsity of the BOW input?
Typically people train autoencoders on the sentence level and not for entire articles. That may be why you are having trouble training your model.