Source: Deep Learning on Medium
At some point in the past, you have probably received spam in the mail that was intended to look like a handwritten letter. The reality, of course, is that it was just printed in a typeface designed to look like handwriting. And that was most likely obvious to you as soon as you looked at it — the letters are just too consistent to be real handwriting. But that doesn’t mean it isn’t possible for a robot to convincingly reproduce handwriting, which is exactly what this algorithm enables.
This algorithm was developed by Brown University undergraduate student Atsunobu Kotani, with help from a faculty member. The purpose of the algorithm is to allow a robot to write letters, numbers, and anything else using pen strokes similar to how people write. That’s a surprisingly difficult goal to achieve, because computers don’t “see” the same way you do. When you look at the letter “A” you see three distinct lines and can easily imagine how to reproduce them with three strokes. A computer just sees a collection of pixels. Kotani’s algorithm converts those pixels into strokes so it can write the letter like you would.
Kotani used deep learning networks to analyze characters, determine the most likely series of strokes needed to reproduce them, and then tell a robot which movements to make to do so. Because the algorithm doesn’t have to actually understand the meaning behind the characters, it can work with any language. For instance, it was able to write “hello” in ten different languages that use completely different character sets. That same methodology can also be applied to any kind of line drawing, like a simple sketch. It might seem trivial, but this algorithm could help robots better interact with people in the future.