Interesting. Tough to evaluate quality, very subjective. Still, seems like quite a bit of progress have been made since Hinton's networks dreaming of digits in 2006 - http://www.cs.toronto.edu/~hinton/digits.html
I remember doing something similar a way back, only just using nearest neighbor search on relatively large dataset (1m of human faces with different scales/etc, or sounds in the MP3 voice recordings), rather than passing things through up-and-down the neuro-net. The result was very similar. I wonder if one can get a better result with the dogs as well, just by using nearest neighbor. A good baseline...
I remember doing something similar a way back, only just using nearest neighbor search on relatively large dataset (1m of human faces with different scales/etc, or sounds in the MP3 voice recordings), rather than passing things through up-and-down the neuro-net. The result was very similar. I wonder if one can get a better result with the dogs as well, just by using nearest neighbor. A good baseline...