CODEWITHSUNDEEP
matlabneural-networktraining-dataperceptron
Drakhlur
Drakhlur

Reputation: 27

MATLAB Perceptron from scratch - OR function

my first time asking a question.

I'm teaching myself neural networks and am currently trying to program a perceptron algorithm. I want to train it for the OR function, but it isn't working. I have absolutely no idea regarding what I am doing wrong, and there are no solutions on the internet that don't use a toolbox.

input = [0 0; 0 1; 1 0; 1 1]%input vector
num_in = 4;% number of iterations
desired_out = [0;1;1;1] %desired output
bias = -1; %bias
w=zeros(2,1); %weight vector, initially zero
iterations = 100; % number of iterations to go through

for i = 1:iterations
     out = zeros(4,1);
     for j = 1:num_in %go per row of x
          y = bias+input(j,1)*w(1,1)+input(j,2)*w(2,1) %sum
          if(out(j,1)~=desired_out(j,1))  % modify weights and bias if mismatch exists
             bias = bias+desired_out(j,1);
             w(1,1) =w(1,1)+input(j,1)*desired_out(j,1);
             w(2,1) = w(2,1)+input(j,2)*desired_out(j,1);
          end
     end
end
out %print the output

Upvotes: 1

Views: 1263

Answers (1)

Autonomous
Autonomous

Reputation: 9075

I don't know which perceptron algorithm you are following but I think the one on Wikipedia is what you are trying to implement.

  1. It is better to incorporate bias into weights i.e. w will be 3x1 and you have to append a column of ones at the end to your input features. This will allow you implement wx+b using matrix multiplication i.e. in vectorized fashion.
  2. You are not updating out. You should have added the following line: out(j,1) = y > 0;
  3. Why do you put this condition: if(out(j,1)~=desired_out(j,1))? It is not mentioned on Wikipedia. Anyway, if you want to update only on mistakes, then you have to update differently on mistakes done on positive and negative samples. See this.
  4. Doing input(j,1)*desired_out(j,1) is wrong. According to Wikipedia, it should be (desired_out(j,1)-out(j,1)).

The corrected code is as follows:

input = [0 0 1; 0 1 1; 1 0 1; 1 1 1] % input vector
num_in = 4; % number of samples
desired_out = [0;1;1;1] % desired output
w=zeros(3,1); % weight vector, initially zero
iterations = 100; % number of iterations to go through

for i = 1:iterations
   out = zeros(4,1);
   for j = 1:num_in % go per row of x
      y = input(j,1)*w(1,1)+input(j,2)*w(2,1)+w(3,1); % sum
      out(j,1) = y>0;
      w(1,1) =w(1,1)+input(j,1)*(desired_out(j,1)-out(j,1));
      w(2,1) = w(2,1)+input(j,2)*(desired_out(j,1)-out(j,1));
      w(3,1) = w(3,1)+input(j,3)*(desired_out(j,1)-out(j,1));
   end
end
out %print the output

This could be vectorized further by using matrix multiplications instead of for loops, but I will leave that up to you.

Upvotes: 1

Related Questions