This AI experiment generates realistic handwriting: Bring on the identity stealing robots

This isn’t new – it’s from 2013 – but it hasn’t got nearly enough attention. Alex Graves at Toronto University built this handwriting generation demo that uses recurrent neural networks to produce convincing cursive script.

Here’s the explanation of how it works from the abstract of Graves’ paper on the project:

This paper shows how Long Short-term Memory recurrent neural networks can be used to generate complex sequences with long-range structure, simply by predicting one data point at a time.

The approach is demonstrated for text (where the data are discrete) and online handwriting (where the data are real-valued). It is then extended to handwriting synthesis by allowing the network to condition its predictions on a text sequence. The resulting system is able to generate highly realistic cursive handwriting in a wide variety of styles.

The demo allows you to select from some predefined samples or have a new one generated. You can also mess with the variables a little.

Here’s how it did with the ‘Tears in the rain’ speech from Bladerunner:

Okay, so the neural network needs to improve its penmanship, but it’s still an incredible bit of work. The fact that it repeats every line to give you multiple options reminds me of this:

And if you want to be alarmist, it’s a good step towards evil robots stealing your identity. Woah, dude!

Still, I’m pretty sure I spoke to the real Bill from Bill & Ted recently.

➤ Recurrent neural network handwriting demo [Alex Graves/Toronto University]

Read next: You can now download a dataset of 1.65 billion Reddit comments: Beware the Redditor AI

Feat image credit: Warner Bros.