You can’t have a real world samples for every problem that could arise. But I think I have a start of a ?solution for this.
Using Generative Adversarial Networks or GAN’s. The kind of network that is trained to tell the difference between fake and real and generate real looking fakes.
It should be possible to create ’photo realistic’ fake problems. Then train a ?separate model on these imagination examples.
This way you extend real world example problem model with a imaginative based model which would be trained by a ?large number of real looking problems.
With a GAN network it would also be possible for a human engineer to do paint assist problems. That is. You would roughly paint the start of a problem and the network would iterate a photo realistic video of the problem.
One usage area ?could be self driving cars.
// Per Lindholm