Perceptron problem

[Arucard]

First few levels were no problem, but I'm having issues understanding how to optimize the first perceptron task.

I assume the idea is to "train" up the accuraccy until it's at an acceptable level, but it seems to max out at around 20% with a lot of errors. I thought using two in conjunction would work, but they don't share accuracy so it just takes more time and makes even more mistakes.

Am I supposed to accept the low score and come back to optimize it later, or am I misunderstanding something here?

[The Cakemeister]

Perceptrons max out at 20% each, it's (largely) something you just have to accept, and then ignore completely when you get genetic algorithms, which go to 9%.

That being said, using multiple in conjunction will boost accuracy, by using two layers you can get ~95%.
Make three perceptrons, each designed to output to one shape, and have them send the other shapes to the appropriate perceptron. (that is to say, the circle one outputs circle to the circle output, and sends squares and triangles to the perceptrons outputting squares and triangles)
Then you simply use a perceptron to filter the input into the three output perceptrons.

It's important to note, however, that it only needs 40% accuracy from you, so you don't really need to go through that trouble, although given how nodes work, doing so will actually help out (slightly) in the long run.
#33 This is actually how you need to do #33, although it only refines for 1 color and trashes the rest, but if you set up your pretrained genetic algo to be a double layer you can actually just run the red output straight to the output and the g/b output to the trash and be done with it.

[Arucard]

Ok, so nothing out of the ordinary is happening, I just got kind of confused with all the errors, but as long as it's normal. Just have to pound down my OCD and move on I guess.

I'm still way in the beginner parts, so I'll play around a bit and see what I can figure out, thanks for the help. I may need more later ;)

[Marcus Germanicus]

my perceptrons color of task #32 increase the error from 75% to 100% .. is that a bug or is it supposed to be a special challenge? :D

[GospodinNoob]

Originally posted by Just Marc:

my perceptrons color of task #32 increase the error from 75% to 100% .. is that a bug or is it supposed to be a special challenge? :D

Hi!

Try to use Gradient descent to training perceptrons!

[Arucard]

I still haven't gotten that far, but the number/percentage is a little confusing to me.
I saw in some screenshots of an earlier version walkthrough where it actually says error or something similar there. So it is still showing percentage of errors, correct? It's just a bit confusing for me, when the number turns green I assume it means 20% success, but that really means 80%?

[GospodinNoob]

Originally posted by Arucard:

I still haven't gotten that far, but the number/percentage is a little confusing to me.
I saw in some screenshots of an earlier version walkthrough where it actually says error or something similar there. So it is still showing percentage of errors, correct? It's just a bit confusing for me, when the number turns green I assume it means 20% success, but that really means 80%?

The percentage is an "error rate", so 20% means that 80% of elements will be recognized correctly

[Kaws]

A way to use the perception filter to remove errors completely would be to have it produce a %100 error. You basically end up using them as a NOT filter. It then takes 3 steps to properly sort. 1 channel you end up having to recycle through the filters.