Neural networks take time to train, and generate. The training is very repetitive, often adversarial, and the outcomes are different, and unstructured. However, the output is also undeniably some of the most intelligent machine behaviour we see.
What would be interesting is if the activity of training neural networks could be used as a medium of exchange. Think of mining in btc. In btc, those hashes go nowhere and do nothing. Yes, neural network training isn't exactly fungible, due to various tasks being non fungible, but that does not mean it isn't extremely useful.
I think iota could step in somewhere here. I am NOT proposing that network based pow be abandoned in favour of some sort of arcane neural network training task with no clear parameters. However, neural network training time is a valuable commodity, especially on modular neural networks. if we had AIs that lived on our home laptops, downloading the latest module to make them more intelligent would be absolute must have. Google would pay for that training data, as would apple and microsoft.
I am not sure exactly how to harness this into the iota framework, which is why I am coming for a round of discussion and ideas. Perhaps this could become a project, perhaps it could become a core part of the iota protocol as some sort of proposal. Perhaps it could end up being a side project. Who knows. What I do know is that the neural network training is one of the most computation intensive tasks, and it takes money and compute time and its raw product is some of the most valuable stuff we can get our hands on in terms of smart tech.