Computer Science Neural Network

Computer Science Neural Network

Rezwan
rezbhai@gmail.com

CSE 523: NEURAL NETWORK

ASSIGNMENT: 1

  1. Firstly I input all p0 to p9, where p3 is representing the numeric number 3.

  2. Then I input the manually 2 bit flipped errors, pe0-pe9, where pe3 represents the manually 2 bit flipped error of the number 3.

  3. Now I start with p0, p1 and p2 and see if the patterns are recognized. We use the code for pseudo inverse pinv(P)[1] , where P is the combination of the matrix we have by making column first and concatenating into P. So we calculate weight as, w = P*pinv(P).   Notice I have to put (:) after my matrixes each time in several points as they need to be made in column (30X1) form to work. Testing,reshaping to (6X5) and viewing of p0 after getting the w is done with one single command:

      >> imshow(reshape((hardlims(w*p0(:))),6,5));

      Using this command I can quickly check my results just by changing p0 to p9 or to pe4 or to pe2.

  4. Testing with the 2 bit flipped errors shows no problem, the neural network of 6X5 grid with three classes saved has learned to give 100% efficient performance by 2 bit flipped error.

  5. Now I directly move to using P having p0 to p9 in it due to time constraint and test whether the network learns the 10 digits.

      The results were 100%.

      The neural network of 6X5 grid with ten (10) classes saved has learned to give 100% efficient performance by 2 bit flip error.

  6. Now we add ten more classes to the network to see the tolerance of the network, can it save up to twenty (20) different classes and give perfect result if a 2 bit flip error is introduced.

  7. Input c0-c9, (A-J), and ce0-ce9, where ce0 is the intentionally 2 bit error character “A”

  8. Finding P and w for the 20 classes (p0-p9 and c0-c9)

  9. Testing for (p0-p9 and c0-c9). Test 100% successful.

  10. Testing for(pe0-pe9 and ce0-ce9). FAILURE

      Other than the numbers “0” and “8” all had errors....