Hmmm... while I like the concept, I'm a little bit unsure about how the data is encoded. My understanding is that classic Mario games were all designed as repeated "chunks" of tiles, hence patterns of blocks that recur throughout the game.
This is non-trivial -- part of the learning curve of Mario is the fact that you become more fluent/competent in these patterns as you play, and then adapt your strategy based on the different obstacles before and after.
A machine learning algorithm based on the individual tiles may spot the repeating patterns and implement them incidentally, but not necessarily... in which case the levels would not be "Mario" levels.
If you're looking at generating levels at the tile level, pretty much any platformer of a similar vintage would be more appropriate, as most other games had fully hand-crafted levels. But then again, 2D Mario games were always huge because of how quickly the levels could be generated, and other platformers don't provide the same size of training set as Mario.