How to add variation in the results of a neural networks?

Question

I would like to create a neural network that converts text into handwriting for use with a pen plotter. Before I start on this project, I'd like to be sure that artificial intelligence is the best way to do this. A problem that I foresee with this approach is a lack of human like variation in the results. For example, the word "dog", when inputted into the network, would be the same every time, assuming I'm not missing something. I am interested if there is any way to vary the output of the network in a realistic way, even when the input is exactly the same. Could I use a second network to make the results more random, but also still look human-like? Any thoughts/ideas would be greatly appreciated.

Answer 1

I would suggest starting with Generative Adversarial Networks (GAN). They usually are capable of adding some randomness to the output to produce different variants. Moreover, Conditional GANs can generate outcomes regarding the observed condition. Therefore, as you change the condition (as an input to the network) you can get different results.

Some examples of NN for synthetic handwriting generators:

1. https://github.com/sjvasquez/handwriting-synthesis
2. http://blog.otoro.net/2015/12/12/handwriting-generation-demo-in-tensorflow/
3. https://distill.pub/2016/handwriting/