>>12759156I'm only joking. Of course, I actually am biased towards my own work, but I truly don't believe that there is no commercial future for genetic algorithms.
When we're talking about AI optimisation of course we also really talking two different things even for simple neural networks. You have the hyperparameter step: which you can solve with global optimisation. And you have the normal parameter training step usually done with SGD.
I believe that the latter will move towards using some form of sub-gradient instead. The reason for that is that sub-gradient methods scale the best by far so you can push the upper limits of the number of linear parameters you have in a system. The future will be some kind of truncated or stochastic sub-gradient algorithm.
I am not just shilling that, my work is in global optimisation.
Beyond that of course there are many models in AI and I think that in the future we will see more non-linear underlying structures employed in practical problems. In this case it would be difficult to use SGD or sub-gradient on the problem. We will see a return to traditional NLP algorithms or something new entirely.
