Building an unprofessional neural network

[WhiteCube]

Personally I think neural networks are the second dumbest way to write a program, but on the other hand, I do like the simplicity of a perceptron.

Rather than pollute my mind with the statistical gooble-de-gook of training, I thought I would make a few networks manually, just for fun.

I was planning a NAND function, but that was too ambitious, so I'm working on NOT.

x goes in, NOT(x) comes out.

Using a "rectified linear unit" output function, I got it working with 2 inputs, 1 in the middle, and 1 output.

A version that uses the "sigmoid" output function is taking me a bit longer.

 

[WhiteCube]

First attempt.
x = 0, output close to 1.
x = 1, output close to 0.

3 perceptrons using the sigmoid function

  -10    100
x ---( )---\
            \
             ( )--- not x
            /
1 ---( )---/
  100    -5

Ugly. Does anyone see a better way?

 

[WhiteCube]

Oh no, NOT again.
I cheated and used a genetic algorithm to find suitable weights. Don't ask me what's happening with weight a. I got the GA idea from a book, I had no idea they could be used on NNs.

      a       c
    -----(P)-----
  /               \
x                  (R)---
  \   b       d   /
    -----(Q)-----

LAYER 0
 PERCEPTRON P
 a = -91160832560779.5
 PERCEPTRON Q
 b = 3.34041516534408
OUTPUT LAYER
 PERCEPTRON R
 c = 250.8517772713261
 d = -177.0801589214439

Still ugly.

 

[WhiteCube]

     -a     3a
    ----(P)----
  /             \
x                (R)---
  \   a     -a  /
    ----(Q)----

a > 10 gets the job done

 

[John Joe]

Oh no, NOT again.
I cheated and used a genetic algorithm to find suitable weights. Don't ask me what's happening with weight a. I got the GA idea from a book, I had no idea they could be used on NNs.

      a       c
    -----(P)-----
  /               \
x                  (R)---
  \   b       d   /
    -----(Q)-----

LAYER 0
 PERCEPTRON P
 a = -91160832560779.5
 PERCEPTRON Q
 b = 3.34041516534408
OUTPUT LAYER
 PERCEPTRON R
 c = 250.8517772713261
 d = -177.0801589214439

Still ugly.

which book?

 

[WhiteCube]

I don't do adverts. The book only had 2 good ideas in it, I had to wade through 192 pages of waffle to find them. Not something I will recommend.