Millions or billions of iterations to tune some algorithms to output some data which satisfies a success metric.
AI should be renamed "Brute Force Statistical Computing". BFSC. It is what it is.
It's stochastic gradient descent, so slightly better than blind brute force
There are various optimizers... but sure, it's slightly better than brute force. SBTBFSC.
Indeed, and "hyperparameter tuning", i.e. "try every combination but give that an impressive name so people don't raise their eyebrow at us"
There's nothing on machine learning that makes convergence inherently hard. And there's certainly something we are doing wrong that makes it so data-intensive.
I think we need so much data for deep learning because of the sheer number of parameters in these models.
On the order of 100s of billions