Sparse Autoencoder Features for Classifications and Transferability