wshine neural network activity - TheEvergreenStateCollege/upper-division-cs-23-24 GitHub Wiki
-
There is a 1:1 ratio of input nodes and pixels, each input value is the shade of the pixel
-
The red and blue pixels represent what features a node is looking for when determining it's output. I'm unsure what the second image is.
- I think bright blue represents a weight closer to 0
- bright red indicates a number closer 1
- dim colors probably represent numbers with less impact, ex: closer to .5 red or blue indicate whether its above or below a threshold but the colors are dimmed because they the weight is still very close to the threshold? unsure if threshold is the right keyword here since we are talking about edges and not a nodes activation.
-
They are the same. This is because the preprocesser centers our drawing and the only 12 input nodes we can select are on the edges of the image which are always the same shade.
-
I was able almost able to get this, but it was still highly in favor of 7 and 2 was only slightly gray. I'm unsure but I imagine a secondary function/processer might need to handle cases where the output was unclear. a kind of backup plan in case we are not confident in an output.
- I think i get and mostly agree with this analogy. A compiled binary that doesn't need the internet might still use the environment of the system it's run on and have different or unexpected outputs. While i think a trained model given the same input, should always give the same output?