Is Sparse autoencoder still very helpful, or it is being replaced by some other techniques in deep learning?

Question

These days I am not seeing any posts related with Sparse AutoEncoder (SAE). I read one post of Andrew Ng, 2015 about it. I am curious if it still useful or there are some other better methods in deep learning for this

Is Sparse autoencoder still very helpful, or it is being replaced by some other techniques in deep learning?

Answer 1

Take a look to variational autoencoders. They enforce that the activations/latent features follow a normal distribution (0 mean -> sparsity, 1 variance).

Once trained you could sample recostructions from the decoder feeding it with normal samples.