There was an article on either medium or towards data sciences that showed auto encoders are actually pretty good on this. I don't remember the article url, but it may have been this. https://towardsdatascience.com/how-autoencoders-outperform-pca-in-dimensionality-reduction-1ae44c68b42f