Club Amiga de Montreal

home *** CD-ROM | disk | FTP | other *** search

/ Club Amiga de Montreal - CAM / CAM_CD_1.iso / files / 593a.lha / Neuro_v1.0 / Neuro.doc < prev next >

Wrap

Text File | 1991-12-13 | 15.8 KB | 369 lines

--------- | Neuro | --------- A neural network simulator Version 1.0 by Berthold Ruf and Ulrich Wisser Contents: ========= 0. ............. Copyright 1. ............. Introduction 2. ............. Getting started 3. ............. Description of the menu items 3.1 .................. Project 3.2 .................. Action 3.3 .................. Settings 4. ............. Notes to the internal structure of the program 5. ............. Known Bugs 6. ............. Appendix 0. Copyright: ============= This program is shareware and NOT (!) public domain. When copying this program be sure to include the following files: -neuro the program -neuro.doc this documentation -req.library copy this file to your libs: directory Thanks to Bruce & Colin for the excellent library. Their address is: Bruce Dawson, Colin Fox Box, 1215 Daviestreet. Vancouver, B.C., V6E 1N4 -data/Numbers demonstrates, how the net can recognize numbers, works very well -data/Encoder Example of finding an internal binary-like representaion -data/Letters the letters in a 5x5-representation are all quite simular, one pixel is used by many patterns, this example shows the problems for neural networks to learn patterns like these. -data/Pattern05x05.06 very easy to learn, try it yourself! -data/Pattern10x10.04 works in Hopfield-mode quite well IT IS FORBIDDEN TO SPREAD THE PROGRAM SKIPPING ONE OF THESE FILES OR WITH ANY FILE BEING ALTERED! REMEBER THIS PROGRAM IS STILL UNDER COPYRIGHT AND BY THIS REASON PROTECTED BY LAW! THE AUTHORS DON'T GUARANTEE ANY FUNCTION OF THE PROGRAM AND ARE NOT LIABLE FOR ANY DAMAGE RESULTING IN THE USE OF THIS PROGRAM! If you like this program and use it more frequently, please send DM 20 (Europe) or $20 (Oversea) to: Berthold Ruf / Ulrich Wisser Mauerstrasse 44 W-5100 Aachen Germany You will get a free update of the next version, as soon as it is available. If you have any questions concerning our program or suggestions for improving it, please write or email us. Our internet-addresses are: ruf@cip-s01.informatik.rwth-aachen.de wisser@cip-s01.informatik.rwth-aachen.de voice : +49-241-33176 Many thanks to our beta-testers, who have helped to improve our program. Especially AnTu did a great job. He also created the nice icon. 1. Introduction: ================ The program "Neuro" is a neural network simulator. It learns patterns and recognizes them, also if the input is incomplete! It can for example learn the pixel representation of some letters. After the user entered a part of a pattern the network will complete this pattern. Used in the mode "Backpropagation" it finds internal represantations for patterns. So the net is able to "discover" the binary code for example, when it has to learn 8 patterns, but has only 3 Bits (hidden units), to represent them internally (known as the Encoder-problem, see the included example-file "Encoder"). Two types of neural networks are implemented in this version: -Backpropagation -Hopfield In the next version we also plan to realize the Boltzmann-mashine. It is not possible to give a full explanation here, if you are interested in details, please refer to the books listed in the appendix. The backpropagation-algorithm is probably the one which might get the best results. The net is able to learn most of the pattern-association problems very well. The hopfield-net also very often produces good results, though it can't distinguish between a pattern and its reverse partner, so you might sometimes get a reverse pattern as an answer. Unfortunately it has a very small capacity so you can't enter too many patterns. In this case you will get a warning message. The programm offers the opportunity, to explore the possibilities (but also the limitations) of neural networks. 2. Getting started: =================== Before starting you should copy the "req.library" to your libs: directory. The program is started from CLI (or shell) with the command neuro or from the workbench by doubleclicking the Icon. No parameters are necessary. For a first try, select "Load all" in the menu "project" and load the file "data/Numbers". You have loaded now a neural network which has learned to recognize the ten digits (0-9). To see them, select "info" in the menu "project". Now press a key or a mouse button and select the item "question" in the menu "Action". You now will see three grids. The left one is for your input, the middle one shows the internal representation of your input and the right one shows the answer of the network. Click in some squares of the input grid, so that it looks simular to one of the digits as shown in the info. Now click "question" in order to see the answer of the network. If it doesn't look simular to one of the digits you can take this answer again as an input by selecting the gadget "Retake". You also can find out, how certain the network is about its answer by clicking into "Black&White" which then will change to "Grey": Some squares will become grey. The more black a square is, the more sure the network is that this square has to be black. 3. Description of the menu items: ================================= 3.1 Project: ---------------------- -Load: You either can load the patterns only or the patterns plus all parameters and the weights of the matrix (the result of the learning process). When the program is started you either must load something or generate a new pattern file with the included editor (see below). When you select "Load patterns", an ASCII-file containing the descritption of the patterns to be learned is loaded. For a detailed descritpion of the structure of the ASCII-file see chapter 4. The file loaded by "Load all" is a Non-ASCII-file. It contains the patterns, the mode of the net (Backpropagation/Hopfield), the weights and the current values of the parameters, which can be changed in the "Setting/parameter" menu. -Save: You only can save the patterns with this menu item, which is not nessessary since you have generated or changed patterns with the editor. After learning, you should choose "Save all" in order to save the results of the learning process. -Clear: This will reset the complet program. All patterns and learning results will be deleted! -Edit: This is a very simple editor, but it works and it should suffice to create or change patterns. --New: If you want to create a completely new pattern file, select "Edit new" (the old patterns in memory will be forgotten then!). You have to tell the editor, how many rows and columns you want to use for your patterns. Then you can click with the mouse in these squares which you want to have black in your pattern. When finished, select "Take" in order to use this pattern. The grid should become empty and you can generate the next one. You can page through the patterns forward and backwards using the "Prev" and "Next" gadgets. If you want to remove a pattern from the list choose "Remove". To clear the grid choose "Clear". A maximum of 30 patterns can be edited. --Old: You can change your edited patterns later or a given pattern file by selecting "Edit old", which always refers to the current patterns in memory. -Info: Gives information about free memory, size, number, parameters and layout of the patterns. When the mode is set to backpropagation it also displays the learning error. The net has performed its task very well when the error is nearby zero. It can take a few seconds to calculate it, so don't panic! -About: To see the beautiful greeting-picture again. -Quit: should be no problem! 3.2 Action: ----------- -Learn: Selecting "learn" starts either the backpropagation- or the hopfield-learning algorithm, depending on your choice of the mode. If the learning-process is normally completed, you'll get the message "Learning finished!!!". If you have choosen the backpropagation algorithm, you have to enter the number of learning iterations, that means how often a (randomly choosen) pattern is presented to the net in order to learn it. In most cases you will need 500 to 20000 iterations to get good results. It can sometimes take a few hours (depending on the size of the patterns) to finish. A display shows, how many iterations are still to do, it is refreshed always after 10 iterations (in order to save time). By striking any key a requester will apear and ask you, if you really want to abort. Select Abort to abort or continue to (guess what) continue. The results up to this moment won't be lost!. In Hopfield mode for a pattern file it's only necessary to call this menu item once and it doesn't make sense here to call "Learn" several times. -Question: After selecting "question", a window will appear, displaying the mode (of the network) in the first line . Below you see three grids: the left one is for your input, the right one is used by the network to display the answer of the network to your question and the middle one displays the hidden units of the network (not in the Hopfield-mode), which show the internal represantation of the current pattern. Click with the left mouse button in those squares of the input grid, you want to have black. After clicking the "question"-gadget the network will begin its search. In Hopfield-mode you can see on the left side the current energy-value. Here it can take a while until you see the answer. "Clear" will clear all three grids. "Re-Take" (in Backpropagation-mode) will take the current output again as a input and automatically start the query. "Black&White" / "Grey" is only available in the Backpropagation-mode. The hidden units and the output units contain internally values between 0.0 and 1.0. In "Black&White"-mode all values higher than 0.5 are displayed black the other values are displayed white. In "Grey"-mode the values from 0.0 to 0.1 are displayed white, the values from 0.1 to 0.2 in the brightest grey and so on. So in this mode you get more information espacially about how certain the net is about its answer. "Go-On" (only in Hopfield-mode) : If the network has found a solution which does not correspond to one of the given patterns, it is possible to "unlearn" this pattern by clicking "Go-On". This will reduce the probability of finding this solution. You should not use this option too often. The efficiency of this unlearning-procedure is given by the parameter Mu (see below). 3.3 Settings: ------------- -Parameter: Here you can set several parameters for the current selected mode. For all parameters there is a default-value, which will be displayed here and used by the algorithm unless you change the value. When entering erroneous values outside of the allowed interval the screen will flash and the entered value will be replaced by the default value. --Backpropagation: Eta determines the learning rate and must have a value between 0 and 1. The bigger Eta, the faster the net will learn. But with a value of Eta nearby 1, there will the danger, that the net won't learn all patterns. It's sometimes a good strategy to begin with large values and then to reduce Eta down to a very small value. Theta determines the bias of each unit (see literature) and should have small positive or negative values. The number of hidden units tells the net, how many units it can use to find an internal representation of the patterns. As a default value there is one hidden unit for one pattern. --Hopfield: Theta: As in Backpropagation. Mu: Must be between 0 and 1 and influences the unlearning-procedure invoced by "Go-On" in the question-requester. The current output pattern will be completely deleted, if Mu=1.0. Unfortunately it also will reduce the probability of finding the other patterns. With Mu=0.0 "Go-On" will cause no changes in the weights. In this case the net only will continue its search. -Mode: Here you can set the program to one of the two implemented network types (boltzmann will be realized in our next version). The behaviour of the complete program depends on your choice in this menu item. -Status: This is only for specialists of interest. It displays the weight-matrix in a very (!) simple way. In our next version we plan to improve this display. You see a lot of "+" and "-", where "+" stands for exhibitory connections (weight >= 0.0) , "-" for inhibitory connections (weight < 0.0). -Random-Gen.: The random number generator is always initialized by a special variable called Seed. In order to change the random aspects of the networks (e.g. when initializing the weights with random values) you can change the Seed-value. Seed must be in the range of 0..65536. -Colour: If you like it, you can change the colours. The last eleven colours are used for the grey-scaled representation of the answers in backpropagation-mode and should change continuously (use spread). -Close WB: Closes the workbench, if there are no open windows with running programs. The workbench will open again when you quit the program. This option may speed up the program a bit. 4. Notes to the internal structure of the program ================================================= The pattern file is an ASCII file which has the following structure: 1.line : Number of patterns 2.line : Number of columns 3.line : Number of rows Then in each line follows one line of the pattern containing only '.' and '*'. The patterns are separated one blank line. Even the last Pattern needs + this blank line. If you have, for example, two patterns with 7 rows and 8 columns, the pattern file looks like this: 2 8 7 ...**... ..***... .****... ...**... ...**... ...**... ...**... .******. ..****.. .**..**. .....**. ....**.. ...**... ..**.... .**..... .******. (eof) The name-convention of the pattern-files is: "Pattern" + #Rows + "x" + #Columns + "." + #Patterns. So "Pattern10x08.15" stands for a pattern file with 15 patterns, having 10 rows and 8 columns. The files created by "Save all" can't be edited by an editor. 5. Known Bugs: ============== When entering erroneous parameters into a gadget, the cursor in this gadget will not disappear. If you should find more bugs, we'd be happy to get a note from you! 6. Appendix: ============ Interesting books: J. McClelland, D. Rumelhart: Parallel Distributed Processing. Explorations in the Microstructure of Cognition. The MIT Press 1986. E. Schoeneburg. Neuronale Netzwerke. Markt und Technik 1990. (in german)