Download ЛЖЖЛ - Cognitive Systems

Transcript
28
CHAPTER 3. NEURAL NETWORK TERMINOLOGY
3.5 An Example of a simple Network
Figure 3.4: Example network of the letter classier
This paragraph describes a simple example network, a neural network classier for capital
letters in a 5x7 matrix, which is ready for use with the SNNS simulator. Note that this is
a toy example which is not suitable for real character recognition.
Network-Files: letters untrained.net, letters.net (trained)
Pattern-File: letters.pat
The network in gure 3.4 is a feed-forward net with three layers of units (two layers of
weights) which can recognize capital letters. The input is a 5x7 matrix, where one unit
is assigned to each pixel of the matrix. An activation of +1:0 corresponds to \pixel set",
while an activation value of 0:0 corresponds to \pixel not set". The output of the network
consists of exactly one unit for each capital letter of the alphabet.
The following activation function and output function are used by default:
Activation function: Act logistic
Output function: Out identity
The net has one input layer (5x7 units), one hidden layer (10 units) and one output layer
(26 units named 'A' . ..'Z'). The total of (35 10 + 10 26) = 610 connections form the
distributed memory of the classier.
On presentation of a pattern that resembles the uppercase letter \A", the net produces
as output a rating of which letters are probable.