I’ve been having some frustrations. Nothing terrible, but many are aggravations. In the interests of maintaining some balance, because I really am in a good place for all my aggravation, I want to mention something really important to me, a positive thing, about engineering.
I can just test it.
I don’t have to take anyone’s word for biscuits. I can just test it. Some theory floats around that sounds like pure nonsense? Boom. Test it.
In deep learning, there’s a point that borders on aphorism: validation data must be distinct from training data. Most references state this as a natural law on par with ‘gravity points down,’ and the shear blind willfulness of it drives me wild.
So I tested it. It’s running now. Took about ten minutes to fork my data appropriately, or not fork it rather, and set the thing to run.
Don’t misunderstand. They’re probably right, and the forked data will probably work better than the other. But I don’t have to believe that at all. I can just test it. There are matters to be taken on faith, including literal faith, but also matters by time and priority I simply can’t duplicate. Virology is the current-events one, but there are plenty of big issues I can’t test. Maybe I could have had I spent the last twenty years learning something different, but I didn’t.
But in engineering, I can to a large extent. So when I’m reading something and start arguing, loudly, with a book, I can throw the idiot to the wind and run my own simulations. And that is not everything, but it is something.