Still in awe of the artist who came up with that one.
It's not the same, but it's left-right symmetrical instead and really neat.
There's probably a way to adapt the training process to create black and white images directly, e.g. by evaluating the classifiers on the thresholded images, but passing the gradients through to the underlying continuously-valued images.
https://upload.wikimedia.org/wikipedia/commons/d/d5/SUN_micr...
entrée / sortie (entrance / exit)
bigram<->character ambigrams are also feasible, but the search space gets pretty big.
Aside, I'm really psyched at the idea that the word "dream" has a chance to become an established term of art here. Do androids dream of electric sheep? Well, it depends on the model you've trained them with, obviously.
It's one of those things I learned the basics of several years ago and have now forgotten (which frustrates me, but is understandable since I don't use machine learning at work).
Say your hands were 2d, so the backs looked the same as the palms. Now hold your hands so that you're looking at the palms and the fingers are pointing up. Rotate your left hand 180° along the y axis, and it would look like your right.
https://www.google.com/search?q=scott+kim+inversions&tbm=isc...
The book is out of print, but you can find used copies. It has a Foreword by Douglas Hofstadter and a Backword by Jef Raskin.
I bought an autographed copy at a computer show in San Francisco on March 21, 1982. The way Scott autographed the books was by inventing and drawing your own personalized inversion on the spot! Mine reads with my first name right side up and my last name if you turn it upside down.
One clever thing I just noticed: my first name has "ae" in it and my last name has "ea", so that made a natural way to link the inverted names together.
Needless to say, this is one of my most treasured books. Thanks Scott!
"The Ambigram deck itself seems impossible at first, and in fact, it would have been impossible just a few years ago. No human being could have designed it alone, even if they possessed several lifetimes to work on the problem."
which is exactly what Cerulean's page refutes - it's a hand-designed 7x7 grid where each word transforms into each of the other words.