Feed it two inputs (e.g. chance of rain and wind speed; this is one of the examples in the demo) and it learns to answer a yes/no question like "bring an umbrella?" It's a one-neuron binary classifier with three learned parameters: two weights and a bias. Those three numbers map directly to Red, Green, and Blue. Save the model: you get a 1x1 PNG. Load the pixel: you get your classifier back. The color is the model.
Has anyone done this for larger neural nets? Is there a way to extract some kind of pattern or is the image just noise no matter how you construct it? I'd be curious to see something like that
This was just a fun thought experiment.
Feed it two inputs (e.g. chance of rain and wind speed; this is one of the examples in the demo) and it learns to answer a yes/no question like "bring an umbrella?" It's a one-neuron binary classifier with three learned parameters: two weights and a bias. Those three numbers map directly to Red, Green, and Blue. Save the model: you get a 1x1 PNG. Load the pixel: you get your classifier back. The color is the model.
It is fun. The trainer seems to get stuck in local minima though.
Thanks; does this fix it? https://github.com/dvelton/ai-pixel/commit/ba04e1c649e0ec99b...
Tried again and no. To reproduce create a corridor e.g.
Line can get jammed in non vertical positon.Increased default epochs to 2000 instead of 500, I think that fixed it but let me know if you still run into this.
This is super fun! Very elegant idea
Has anyone done this for larger neural nets? Is there a way to extract some kind of pattern or is the image just noise no matter how you construct it? I'd be curious to see something like that
Not sure what happens if you start scaling this, but curious too. Would love to see someone who actually knows ML try it.