Results (
Thai) 1:
[Copy]Copied!
I. Introduction to Neural Networks The ability to recognize and associate character patterns makes neural nets useful in an area such as handwriting analysis. We will illustrate this with an example in which we train a network to learn a pattern for each of the 26 letters of the alphabet. To keep the number of inputs to a reasonable size, the characters will be upper case and defined by a 5x5 array of dots. A more practical size array for character display is 9x7, but this would necessitate 63 rather than 25 inputs. An alternative problem, to be illustrated later with unsupervised learning, would be to train the network to learn several patterns for each character (e.g. different font styles) and fewer characters overall (to reduce the number of patterns to be learned).For example, the letter "A" can represented by the 5x5 dot array below - each position highlighted is shown graphically by the pound sign (#) for which the corresponding input value is +1. The non-highlighted positions, shown by a dot (.), have values 0 for binary or -1 for bipolar. Character Inputs for "A" (binary display shown)The string of bits, from top to bottom row and left to right column becomes '0010001010100011111110001'corresponding to neural net inputs x1 to x25. Each letter will of course have a unique input pattern. We will define 26 separate outputs, one identifying each letter in order A to Z. Thus, if the pattern above for "A" is presented, the binary target output string is '10000000000000000000000000' (t1 to t26). Likewise, for "B" only the second output is high and the rest are 0's, etc.The objective is to train our network to learn the 26 character patterns. If learning is successful, the network can be tested by entering input patterns "close to" a learned character, and see if the net can associate with that character.0MATLAB [1,2] is a widely used matrix based equation solving program. A MATLAB program was written to apply Perceptron to train a neural network for this requirement. The output when the program is run is shown below. For only five training iterations (epochs) most of the character patterns have already been associated with the appropriate output line. All 26 alphabetic characters are learned after 17 epochs, resulting in a unit 26x26 matrix displayed. Note the algorithm is applied with a bipolar threshold function, after which the outputs are converted to binary for convenience in presentation.
Being translated, please wait..
