Brain Damage

I finally finished debugging my neural network code!

Yaay.

As it turns out, there were problems. A little one here, a tiny one there, and those little tiny itty bitty mistakes made for simulated computer brains that were... miswired. Twitchy.

Damaged.

(Solution space of a brain-damaged neural network.)

They're certainly interesting, but I need a solid, correctly-functioning neural network to be able to continue with my... experiments...

MWHAHAHAHA!!!
*cough*

(Another damaged neural network.)

I made very small networks that I could test by working through the calculations by hand as well as having the computer do them. Slowly, bit by bit, the mistakes got corrected...

... aaaaand, where the hell'd the "interesting" go?? Must still be a few bugs left...

Fix two bugs, accidentally introduce a brand new bug.

Suddenly, the brain fractures in a totally different way.

I kinda wish now that I'd saved a copy of this particular version of the malfunctioning C code. It would have made a decent abstract-art generator.

Finally. The code checks out. Its answers match the ones on paper.

I increase the neuron count and run the program:

(First image generated by the properly-functioning neural network code.)

I'm discovering that initial severity of randomness that I start with in the simulated brain needs to be changed, depending on the number of neurons I tell the brain it's going to have.

I increase the number of neurons again, and I start getting more interesting results.

The system is starting to bog down a bit since the simulated brain is getting larger exponentially. I shrink the image size so I'm not throwing as many pixels at it.

  

  

These l'il guys were made with neural-networks large enough to keep the computer busy for several minutes calculating them.

(That's not bad when you consider that the brain has to be run on every single dot in the image. That's thousands to millions, depending on the image size.)

(It took minutes to draw, but each individual pixel only took
about seven one-thousandths of a second to calculate.)