Deep Learning

Deep Learning is a branch of Artificial Intelligence, which is a nonsensical name. A better name would be computers-doing-repetitive-but-useful-math, but that’s not as exciting and won’t get the attendees at a convention. Coincidentally, any time someone tells you the computers are going to take your job through AI, mentally replace AI with ‘addition’ and see if the FUD still works. If not, don’t worry about it.

Anyway, DL used to be called big data, and while they’re not exactly the same, they’re closely related. The computer does a lot of math very quickly, but it’s an awful lot of math so it takes a while. For the math to be useful, it has to do an awful lot. What the computer is doing the math on is the data, and that ‘awful lot’ referenced above means the computer needs a lot of data, ie big data.

With me?

So to train a DL network, I need a lot of data, and that means I spend hours a day waving targets over detectors. It is boring.

It’s boring but necessary in the way getting fit requires eight hours of sleep a night and eating a healthy diet. No one wants to do that.

Lifter: Coach Matt, I want to get jacked.

Coach Matt: Sounds good. Start with 8 hours of sleep a night and a healthy diet.

L: That’s boring. I see the huge guys in the gym lifting. How do I do that!?

CM: Patience, 8hrs of sleep a night, and a healthy diet.


Lifter leaves.

L: Coach Bob, how do I get jacked?

CB: Benchpress. Chest day, benchpress. Shoulder day, benchpress. Leg day, benchpress. Ab day, benchpress. Back day? No one cares about your back. Anyone saying anything to your back should come around and say it to your face. And if they’re staring at your face, you should work on your front. Benchpress.

L: There is no way this can go wrong.

Meanwhile, like a loser, I’m trying sleep right, eat right, and stretching.

The lab equivalent of this is taking data, and it’s boring. Maybe I should wave a target over a detector for 8 hours and get some data!? Fun? No. Not fun.

The trick IT companies use is that they have a network of engineers doing various jobs, so one set of engineers creates the data collection system, another develops and implements the network-trainer algorithm (algorithm means ‘specific process done by computers’), a third examines the trained network and so forth. This cycle is sometimes called the product development triangle, and once you go around it five or six times, iterating and improving, it gets pretty good. Then the DL algo works well, some IT company feeds it huge data, and they try to take over the world.

Until their CEO runs off with all the money.

Come to think of it, maybe the ethical choice is just the benchpress. You don’t accomplish ****, but no one makes enough money to commit fraud.

My data is processing. I’m really bored.

Leave a Reply