Of course that is remotely possible and would be trumpeted as proof of Ai mental brilliance where actually it is simply an example of a blind squirrel running into a nut.
What I’m getting at is that we’re not sure if it’s comparing the shape, the color, the overall luminance, or what when it’s deciding what’s a cirrus cloud or a cumulus cloud. Much like it’s often hard to actually articulate how we decide what the differences are.
Of course we can see the values of variables, set watches, and generally see the states of the machine while it’s running, but in terms of how it’s learning, that’s something a little bit different. I mean, we can see weighs and biases, but do we know what they actually represent?
It’s like grilling something correctly. I have had the hardest time trying to actually explain how I can tell when something’s not done, done, or overdone to my kids. I mean, I could put a thermometer in a steak, but that’s not the same thing. It’s something you learn through experience and it’s hard to articulate.
From what I understand, a lot of machine learning is very similar.
Well, the way the current models work, you wouldn’t really get that. Right now, they’re trained by taking a bunch of photos of clouds that are categorized, and then loading them into the model. Then you take some that are not categorized, and let the model decide. Then you correct it when it’s wrong, and confirm when it’s right. It does its thing, and eventually gets pretty good at identifying which type of cloud photo corresponds with each category.
I’m using clouds because the AI course I took a while back used that as its training materials, but in reality most models are a lot more complicated- finding cancer in MRI scans or things like that.
Again, we don’t usually do it as it adds unnecessary complexity, but that doesn’t mean that we can’t. While they’re called hidden layers, they’re not truly hidden from the programmer. Both the book I linked (which is really worthwhile as it has you do everything from scratch, much like another favorite of mine, Data Science from Scratch) as well as this white paper (among many) show how that works. Yann LeCun used to do a stellar presentation on this, but I’m guessing he’s a bit too busy these days to share with us plebes. ML and DL are extremely cool and powerful, but they ain’t magic.