Sketch form explanation of my little deep learning project.

I walk in sipping coffee, see my program sitting at its desk typing on its typewriter. “So, how’s it going? I ask.

“Got it down to about 1.5% error, take a look.” My computer responds, producing an architectural print from a neat pile next to its desk and unrolling it for my review.

“I see… Well there’s a line missing here.”

“I find that there isn’t really a line there.”

“Well, her neck has to continue, even though the original photo looks dark at that point, otherwise there would be nothing supporting her head and blood wouldn’t be flowing to her brain.”

“Uhmm, I don’t know what neck means, or head, or blood, or brain, or flowing. I’m computing convolution matrices to detect and then accentuate lines… like you asked.”

“Right right, that’s why I set you up with parallel convolution paths which are then recombined in various ways. So you can find these hidden lines in those and then recombine them into the product.”

“Uh huh… well, I find that messing with those can lead to wild increases in error, and i know you don’t like that, so… I don’t.”

“That’s disappointing. I may need to break this project up a bit for you. Well, keep working on it in the meantime.”

“I’m running low on memory, so I can’t do any of your new instructions. Anyway, I’ll just keep working on what I’ve been doing.

“Super.” I take another sip of coffee and walk out.

Leave a Reply