Back when we believed deep network pre-training could be achieved using stacked autoencoders trained with group sparsity regularization.

→ View original post on X — @ylecun