You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I'm increasingly confident that the only reason we don't use neural nets for everything ever is that they are inconvenient. It's super easy to take first differences, second differences, and more; the only problem is that it's annoying to take gradients of big neural nets, but that's automated too now, via packages like Kayak. I think I'm going to ask Gary about it and see what he says. Political scientists have their methods of research in which we look at P-values; economists write a formal model, run an experiment, and see if the model's results match the data; computer scientists run computing time experiments and stuff. I would like to do a hybrid, where I run a neural net and generate predictions, then test the predictions by running a survey experiment or something. It seems like a reasonable response to economists.