Larger problem: mnist, images, and speech are probably low hanging fruits with rich features, large datasets, balanced datasets, etc. for which off the shelf packages work very well today.