JavaNNS User Manual
Contents
1. 3 55648 I2 0423398 17 33895 22 6 03206 27 0 01723 32 0 09160 23 1 07894 28 1 77930 33 1 59529 1 57236 9 0 74423 34 0 13875 5 5 30719 10 2 13168 2 34832 20 5 00616 41 I 1 4 41380 6 1 48152 11 2 62748 1 00557 21 0 06430 12 3 93844 17 4 01591 22 76102 27 0 36823 32 0 54661 23 4 15954 28 2 96118 33 3 30219 0 24202 9 1 56077 34 0 20287 5 1 46062 10 79490 96920 20 3 72459 42 1 1397383 6 2 53253 11 2 04922 13969 21 1 81064 12 0 32565 17 64358 22 02883 27 1 05720 32 0 71916 23 1 00499 28 1 10925 33 3 18685 2 12575 9 0 36763 32 43 44 45 46 452 3 97560 44 36 4 04184 47 45 4 88750 44 36 0 14327 48 45 1 02597 44 36 2 15042 49 45 1458579 44 36 1 92163 50 45 2 95134 44 36 0 73749 51 45 4 16732 44 36 0 40437 52 45 1 78256 44 36 1 73887 53 45 3 64449 44 36 4 29717 54 45 0 45205 44 36 99719 55 45 0 46855 44 56 45 1 27961 44 36 0 72057 57 45 1 21679 44 36 16523 58 45 3 44107 44 36 05544 59 45 5 06022 44 36 10170 60 45 0 03838 44 36 4 49324 61 45 3 65178 44 36 0 46651 62 45s 1 17767 44 36 0 65937 63 45 1 52270 44 36 3 14334 64 45 3 00298 44 36 4 02590 65 45 0 56820 44 36 2 17370 66 45 4 13007 44 36 1 46460 67 45 1 66534 44 36 1 40745 68 45 0 60032 44 36 4 23531 69 45 4 31415 44 36 0 28689 70 45 0 54085 44 36 22. 02860 01347 27674 2 17011 4 Se 92706 60368 38773 64082 15332 65062 11836 90943 01833 38939 94756 09881 86233 11771 44175 65889 18638 35289 49606 05007 05268 50213 67009 97460 01853 77810 19606 44547 229 15 63405 09076 25699 36290 20720 19947 20 21 321 20 21 323 20 21 32 20 43881 36689 45891 86340 43783 24280 55429 92937 43438 41481 23237 54990 28834 80876 24993 53220 25429 41775 47295 52796 14214 43160 1 16736 47655 12564 47751 90942 03714 90286 22540 00602 98775 32045 78425 58451 60600 92319 54519 22067 generat network source no of no of no of no of learnin update site de site n inhibi excite outType Longero connect ed at Fri Aug 3 00 25 42 1992 name xor files units 4 connections 5 unit types 2 site types 2 g function Quickprop function Topological Order finition section ame site function E Site_Pi Site_WeightedSum utType Act Logistic Act Logistic Out Identity Out Identity bias st subnet layer act func out func CX gu cm j pR Gore Aet SS Se 0 0 00000 h 0 1 Act Logistic Out Identity ANSSET a
3. 039 i n p 5 722 nn 5 722 2 lt 3 dis E 6 4 443 a B 1 051 1 051 1 0 4 0 0 E 0 10 20 30 40 50 60 70 80 go 100 x ae Learning cycles TIR Panel Lx SA Fr Initializing Updating Learning Pruning Patterns DAE Step 0 SSE 1 4107024669647217 1 016804575920105 Learning function Rprop y 2 0 9440706968307495 0 5550776720046997 0 21857595443725586 0 08159147202968597 0 0370437353849411 0 023498933762311935 0 015498396009206772 0 01296208519488573 Parameters 50 fos max foo a jo Cycles fi 00 Steps fi I Shuffle Init Learn current Figure 1 JavaNNS with XOR network error graph control and log panel 4 5 Analyzing Network For analyzing the network and its performance tools like Analyzer in the Tools menu and Projection in the View menu already familiar to SNNS users can be used For Projection two input units and a hidden or output unit have to be selected in order for the menu item to become enabled The Projection Panel than displays the activation of the hidden or output unit as a function of the two inputs The activation is represented by color so a colored rectangle is obtained Analyzer is used to show output or activation of a unit as a function of other unit s activation or of the input pattern Its usage is similar to the Analyze panel in the SNNS 4 6 Creating a Network Now let s create a network of our own Choose File New to remove the current network from the
4. 79769 71 45 0 00018 44 36 3 29634 B 2 Example 2 SNNS network definition file V3 34 0 18372 5 LLL ALS 17 231 5 17387 28 12 2 77766 17 23 1 95344 28 12 3 04575 17 23 5 96261 28 4 69081 24788 68170 11916 89325 01698 85157 31087 92594 21280 00822 58017 45729 93490 33955 827713 385 12 39376 19092 64443 60025 44890 84431 0 81393 13145 18362 08796 59148 61389 70962 42352 65709 37122 30654 99220 75100 41717 80036 11360 43 43 43 43 43 43 43 43 43 43 43 43 43 43 43 43 43 43 43 43 43 43 43 43 43 43 22 33 22 Sos 223 33 263 15 24533 23107 30420 39609 60419 97236 55796 08897 13199 23126 97409 31309 16526 72110 04974 89166 95486 46879 50408 57915 23345 1 80938 2 66027 69538 60547 74552 98185 76429 11298 59287 65150 89143 63380 22676 17663 77030 19014 85345 N N N KS EN N 42 42 42 42 42 42 42 42 42 42 42 42 42 42 42 42 42 42 42 42 42 42 42 42 Nop ty DO BAY DO BAI Hw
5. Height h Top left position ht h h Unit detail Unit type funknown El Activation function act Logistic Output function Jout identity H Layer number 6 Subnet number b Create Close Figure 5 Create Layers tool nits 14 Connecting selected units is a two step process In the first step the user selects units where the connections are to originate source units and presses the button Mark selected units as pr ou source In the second step the user selects the receiving units 99s Ed Create links MEI C Connect selected units targets and presses the button which is now labeled Connect K h cios source with selected units For auto associative connections it C Auto associative suffices to select the desired units and press the Connect Allow selt connectior selected units button Selected units can be dragged with the mouse in order to Connect change their positions 5 3 Editing Units Close Existing units can be edited by selecting them and then choos ing Unit Properties from the Edit menu or Edit Units in the con text sensitive menu which is accessed by pressing the right mouse button while over a unit An extra window appears displaying all editable unit proper ties like name type activation etc This method allows only for setting the same values for all selected units Alternatively the user can edit values displayed as top and base labels of each unit indivi
6. L BRACKET paramlist R BRACKET REMAP PARAM L BRACKET paramlist R BRACKET zz NUMBER paramlist NUMBER pattern pattern list pattern pattern start pattern body pattern class pattern start pattern body pattern class actual dim zz NUMBER pattern body NUMBER zz NUMBER NAME 29 30 Appendix B Example Netvvork File The lines in the connection definition section have been truncated to 80 characters per line for printing purposes B 1 Example 1 network name klass source files no of units 71 no of connections 610 no of unit types 0 no of site types 0 learning function Std Backpropagation update function Topological Order unit default section act bias st subnet layer act func out func Arr VERSER 2 2272259823 23953552237 EE 0 00000 0 00000 h 0 1 Act_Logistic Out_Identity SHER tas SS S 2 259 e 2052257 2222292755522 225553 225 unit definition section no typeName unitName act bias st position act func out func sites A e si e ai Y E EA l es ass SA AAA I va Ge n i e e S I A se 1 ull 1 00000 0 00000 i da by D 2 ul2 0 00000 0 00000 i 2 deU 9 u13 0 00000 0 00000 i is LO 4 ul 0 00000 0 00000 i 4 1 0 5 ul5 00000 0 00000 Di 0 6 u21 00000 0 00000 i lom p O 7 u22 00000 0 00000 i 2 20 8 u23 0 00000 0 00000 a SQ 30 9 u24 00000 0 00000 i 4 2 0 0 u25 00000 0 00000 T Dips 22 0 1 u31 00000 0 00000 i Il 3 0 2 u32 0 00
7. License Any attempt otherwise to copy modify sublicense distribute or transfer JavaNNS is void and will automatically terminate your rights to use JavaNNS under this License However parties who have received copies or rights to use copies from you under this License will not have their licenses terminated so long as such parties remain in full compliance 6 By copyine distributing or modifying JavaNNS or any work based on javaNNS you indicate your acceptance of this license to do so and all its terms and conditions 7 Each time you redistribute JavaNNS or any work based on JavaNNS the recipient auto matically receives a license from the original licensor to copy distribute or modify Java NNS subject to these terms and conditions You may not impose any further restric tions on the recipients exercise of the rights granted herein 8 Because JavaNNS is licensed free of charge there is no warranty for it to the extent per mitted by applicable law The copyright holders and or other parties provide JavaNNS as is without warranty of any kind either expressed or implied including but not limited to the implied warranties of merchantability and fitness for a particular purpose The entire risk as to the quality and performance of JavaNNS is with you Should the program prove defective you assume the cost of all necessary servicing repair or correction 9 Innoevent will any copyright holder or any other party who may redistri
8. involved in the SNNS project They are listed in the order in which they joined the SNNS team Table 1 SNNS and JavaNNS project members Andreas Zell Design of the SNNS simulator SNNS project team leader Niels Mache SNNS simulator kernel really the heart of SNNS parallel SNNS kernel on MasPar MP 1216 Tilman Sommer original version of the graphical user interface XGUI with integrated network editor PostScript printing Ralf H bner SNNS simulator 3D graphical user interface user interface development version 2 0 to 3 0 Thomas Korb SNNS network compiler and network description language Nessus Table 1 SNNS and JavaNNS project members Michael Vogt G nter Mamier Michael Schmalzl Kai Uwe Herrmann Artemis Hatzigeorgiou Dietmar Posselt Sven D ring Tobias Soyez Tobias Schreiner Bernward Kett Jens Wieland J rgen Gatter Igor Fischer Fabian Hennecke Christian Bannes Hans Rudolph Radial Basis Functions Together with Giinter Mamier imple mentation of Time Delay Networks Definition of the new pat tern format SNNS visualization and analyzing tools Implementation of the batch execution capability Together with Michael Vogt implementation of the new pattern handling Compilation and continuous update of the user manual Maintenance of the ftp server Bugfixes and installation of external contributions SNNS network creation tool Bignet implementation of Cas cade Correlation and printed
9. item acy Parameters min fio max fo Figure 8 Control Panel Init control Panel 1x Initializing Updating Learning Pruning Patterns Subpattems Learning function Parameters 00 0 35 max foo a Eo Cycles 500 Steps fi I Shuffle Init Learn current Learn all Figure 9 Control Panel Learning During training the error is also written into the log window Also many other useful infor mation about the network are written there on diverse occasions The log window corresponds roughly to the command shell window from which SNNS is started in a Unix system Options and controls for pruning networks are found under the Pruning tab in the Con trol Panel Its contents corresponds mostly to the Pruning window in SNNS However contrary to the SNNS the user does not have to set the learning function to Pruning FeedForward In JavaNNS it is done auto matically and transparently for him her The learning function as set under the Learning tab as well as number of cycles correspond to the data entered in General parameters for Training section of the SNNS Pruning window In JavaNNS prun control Panel 1x Initializing Updating Learning Pruning Patterns Subpattems Method Optimal Brain Surgeon y Maximum error increase 96 fi 0 v Recreate last element Accepted error Fl Refresh display Cycles for retraining 100 v Pr
10. network view apply Arrows at the panel edges are used for moving the projection window in the input space The two buttons on in the top left corner are used for zooming and the buttons in the bottom left corner for adjusting the view resolution Zoom ing can also be performed manually by drag ging a rectangle in the projection area 8 2 Weights Panel The Weights panel presents link weights as col ored rectangles The x axis is used for source units and the y axis for the target units of the links The two buttons at the panel bottom are used for toggling grid and for auto zoom for optimal display As in the projection panel zooming can be performed manually ES Linkweights 10 source E Figure 14 Weights panel 8 3 Analyzer The Analyzer is used to show output or activation ofa unit as a function of other unit s activation or of the input pattern Its usage is similar to the Analyze panel in the SNNS The control buttons are also familiar and have the same function as in the Error Graph Projection and VVeights panel 19 Ed Analyzer 100 120 140 Pattern Number Figure 15 Analyzer panel 20 9 Loading Saving and Printing File loading saving and printing of results is performed through the File menu Whereas Open can be used for loading any type of file network pattern text Save as well as Save as are used only for saving the current network Other file types are
11. on JavaNNS means either JavaNNS or any work con taining JavaNNS or a portion of it either verbatim or with modifications Each licensee is addressed as you 2 You may copy and distribute verbatim copies of the JavaNNS distribution as you receive it in any medium provided that you conspicuously and appropriately publish on each copy an appropriate copyright notice and disclaimer of warranty keep intact all the notices that refer to this License and to the absence of any warranty and give any other recipients of JavaNNS a copy of this license along with JavaNNS 3 You may modify your copy or copies of JavaNNS or any portion of it only for your own use You may not distribute modified copies of JavaNNS You may however distribute your modifications as separate files e g new network or pattern files along with our unmodified JavaNNS software We also encourage users to send changes and improve ments which would benefit many other users to us so that all users may receive these improvements in a later version The restriction not to distribute modified copies is also useful to prevent bug reports from someone else s modifications 4 If you distribute copies of JavaNNS you may not charge anything except the cost for the media and a fair estimate of the costs of computer time or network time directly attribut able to the copying 5 You may not copy modify sublicense distribute or transfer JavaNNS except as expressly provided under this
12. pattern through the network The same panel is also used for selecting the updating function and its parameters to be used in training 16 7 Training and Pruning Networks Training is also performed through the Control Panel In the Initializing tab an ini tialization function and its parameters can be set The Init button also available in the Learning tab performs the initialization Under the Learning tab the user can choose the learning function set its parameters number of learning cycles and update steps and finally perform network initialization and learning The classic Backpropagation equals Std_Backpropagation in SNNS is the default learning function For each learning function default parameters are provided The Learn current button performs train ing vvith the currently selected pattern and Learn all with all patterns from the pat tern set In order to monitor learning progress it is useful to open the Error Graph and or Log window both available from the View menu During learning the error graph displays the error curve The type of the error to be drawn is set through the middle button located on the left edge of the window The arrow buttons near the axes are used for scaling The two buttons in the left bottom corner clear the error graph and toggle grid respectively control Panel Lx Initializing Updating Learning Pruning Patterns Subpatterns Initializing function EET
13. simulator Then choose Create Layers from the Tools menu A window resembling the Bignet tool of SNNS appears Choose width and height 1 unit type Input and click Cre ate to create a new layer For the next layer set height to five and the unit type to Hidden and click Create again Finally create the output layer with the height of one and unit type Output and close the window To connect the created units use Create Connections from the Tools menu Simply choose Connect feed forward and click Connect Doing that you have created a simple feed forward neural network with one input five hidden and one output unit You can now close the Connections window too 4 7 Graphical Network Display You can arrange units on the display manually by clicking and dragging them with the mouse In fact clicking a unit selects it and dragging moves all selected units To deselect a unit press the CTRL Key on the keyboard and click it while still holding the key pressed Using 11 View View Settings tab Units and Links you can choose what to display above and under each unit Make sure that Name is selected as top label Since the units have just been cre ated they are all called noName To change the names choose Names from the Edit menu The top labels turn to text fields Use the mouse to place the caret into each one and enter some names After you have finished press Enter or click in an empty are of the dis play
14. the General tab Minimum values are simply taken to be nega tives of the maximums and for values between the color is interpolated The Slider Weakest visible link is self explanatory and helps keeping the net work view more comprehensible Since more than one network view can be open at the same time Display Settings refer to the currently selected one 5 2 Tools for Creating Networks Networks are created using two tools from the Tools menu both from the Create submenu Layers and Connections They together corre spond to the Bignet tool in SNNS In JavaNNS layer has a different meaning as in the SNNS In JavaNNS layer corresponds to a physical layer of units that is being created When creating layers width and height determine the number of units in horizontal and vertical direction for the layer Top left position is updated automatically but can also be entered manually and controls the layer s position in the display area For all the data width height and coordinates of the top left position the General Units amp Links r Units Top label Maximum expected value Number TO DEDOS i EIOS T ROO ORT Base label OS 1 1 3 Output Links iv Show Show weights Weakest visible link m Urn feo por ebore pore c enm 0 2 4 6 8 10 Maximum expected weight Show directions PE qp or un peg pons neon deg VO OC 0 2 4 6 8 10 OK Preview Default Cancel
15. 000 0 00000 i 25 0 3 u33 00000 0 00000 3 35 033 20 4 u34 0 00000 0 00000 i 4 3 0 5 u35 00000 0 00000 a bus 3 70 6 u41 00000 0 00000 i 1 4 0 7 u42 0 00000 0 00000 i 2 4 0 8 u43 0 00000 0 00000 a Suid 0 9 u44 0 00000 0 00000 i 4 4 0 20 u45 1 00000 0 00000 E 5 4 0 21 u51 1 00000 0 00000 i Tr Deg SO 22 u52 0 00000 0 00000 au yo ty 23 u53 0 00000 0 00000 i 3 Dp 0 24 u54 0 00000 0 00000 i 4 5 0 25 u55 1 00000 0 00000 gh pee EU 26 u6l 1 00000 0 00000 i 1 6 0 27 u62 0 00000 0 00000 i 2 55 0 28 u63 0 00000 0 00000 i 3267 D 29 u64 0 00000 0 00000 4 6 0 30 u65 1 00000 0 00000 i Dr Gr KO 31 u71 1 00000 0 00000 i la fep 32 u72 0 00000 0 00000 ge AR EO 33 u73 0 00000 0 00000 i ak hes 0 34 u74 0 00000 0 00000 4 0 35 u75 1 00000 0 00000 i Bi Tha 0 36 h1 0 99999 0 77763 h 810 0 37 h2 0 19389 2 17683 h 8 1 0 38 h3 1 00000 0 63820 h 8 29 0 39 h4 0 99997 1 39519 h 8 34 0 40 h5 0 00076 0 88637 h 8 4 0 41 h6 1 00000 0 23139 h By 5 0 42 h7 0 94903 0 18078 h 8 6 0 43 h8 0 00000 1 37368 h 8 7 0 44 h9 0 99991 0 82651 h By 8y 0 45 h10 0 00000 1 76282 h 8 9 0 46 A 0 00972 1 66540 o leg 3 240 47 B 0 00072 0 29800 o 25 7 y SO 48 C 0 00007 2 24918 o 3 Le D 49 D 0 02159 5 85148 o 4 1 0 50 E 0 00225 2 33176 o ly 2520 od F 0 00052 1 34881 o 22 y 0 52 G 0 00082 1 92413 o By 25 30 53 H 0 00766 1 82425 o 4 254 0 54 I 0 00038 1 83376 o Ly
16. ACT FUNC COL SEP OUT FUNC COL SEP SITES CUT zz STRING W COL SEP STRING W COL SEP STRING W COL SEP STRING COMMA STRING CUT zz SUBNET SECTION TITLE CUT COMMENT WHITESPACE subnet block subnet header TWO COLUMN LINE EOL COMMENT subnet def TWO COLUMN LINE EOL zz SUBNET COL SEP UNIT NO CUT zz INTEGER VV COL SEP INTEGER COMMA INTEGER CUT UNIT SECTION TITLE CUT COMMENT WHITESPACE unit block iz unit header TEN COLUMN LINE EOL COMMENT unit def TEN COLUMN LINE EOL zz NO COL SEP TYPE NAME COL SEP UNIT NAME COL SEP ACT COL SEP BIAS COL SEP ST COL SEP POSITION COL SEP ACT FUNC COL SEP OUT FUNC COL SEP SITES CUT zz INTEGER VV COL SEP STRING W COL SEP ICOL SEP STRING W COL SEP I COL SEP SFLOAT W COL SEP ICOL SEP SFLOAT W COL SEP I COL SEP STRING W COL SEP I COL SEP INTEGER COMMENT INTEGER COMMENT INTEGER W COL SEP STRING W COL SEP ICOL SEP STRINGW COL SEP ICOL SEP STRING COMMA STRING connection definition section connection section connection block connection header connection def layer definition section layer section layer block layer header layer def 3D translation section translation section translation block translation header translation def time delay section td section td block td header td def 27 CONNECTION SECTION TITLE CUT COMMENT WHITESPACE connection block connection header THREE COLUMN LI
17. Bj 20 55 J 0 00001 0 87552 o Zi Bin 40 56 K 0 01608 2 20737 o 3 3 0 57 L 0 01430 1 28561 o 4 3 0 58 M 0 92158 1 86763 o l 4 0 59 N 0 05265 23452717 o 2 4 0 60 O 0 00024 1 82485 o 35 4 0 61 P 0 00031 0 20401 o 4 4 0 62 Q 0 00025 1 78383 o 1 5 0 63 R 0 00000 1 61928 o 2 7 58 0 64 S 0 00000 1 59970 o 3 5 0 65 T 0 00006 1 67939 o 4 5 0 66 U 0 01808 1 66126 o 1 6 0 67 V 0 00025 1 53883 o Zip 69 0 68 W 0 01146 2 78012 o 25565 0 69 X 0 00082 2 21905 o 4 6 0 70 N 0 00007 2 31156 o lys L0 71 Z 0 00002 2 88812 o 25 Ky D connection definition section target site source weight AA th I SEP I AAA A A A A On ST T 36 p 0950934 G 383328 11r 154422 18840 21 4 59526 SH I 1 0 93678 6 0 68963 11 0 94478 1 06968 21 0 47616 12 2 62854 17 05391 22 0 37275 2 0 12598 32 0 27619 23 1 45917 28 1 97934 33 1 01118 4 39595 9 2 78858 34 0 14939 5 1 80792 10 3 66679 2 53150 20 1 07000 38 1 2 44151 6 0 41693 11 2 12043 1 40761 21 1 83566 12 0 55002 17 2 08524 22 0 63304 2 0 27301 32 2 49952 SS o 00 RN OURAN Oo ns J 0 Ui Jo OU BAI VFO BAO 39 1 5 17748 6 4 45709 11 0 65733 2 26190 21 2 69957 12 1 43420 17 0 33409 22 0 74423 27 1 38010 32 3 08174 23 4 42961 28 1 09858 33 2 09879 1 30835 9 0 79940 34 1 99276 5 2 61433 10 3 56919 00952 20 2 86899 40 1 3 03612 6 0 05247 11 3 20839 4 03382 21
18. Figure 4 Display Settings Units and Links measuring unit is grid size unit which is set in the View Display Settings panel In Unit detail segment of the window the unit type e g Input or Hidden activation func tion of the units Logistic by default output function of the units Identity by default the layer number and the subnet number are set The Connections window provides for creating links connections between units Three different ways are possible for creating links by manually selecting units to connect Connect selected units by automatically connecting the whole network in a feed forward style Connect feed forward and by interconnecting those in the same layer Auto associative In case of feed forward networks shortcut connections links connect ing units form non adjacent layers can optionally be created For auto associative networks self connec tions feedback connections from the output to the input of a same unit can be allowed Except for automatic generation of feed forward con nections the user has to select units to be connected Units are selected using the mouse either by clicking each unit or by clicking the mouse and dragging a rectangle around units to be selected Units are deselected by clicking them while holding the CTRL key pressed A simple click in an empty area in a network view deselects all u Ed create layers x1 r Size amp Position width oo
19. KS TABS STRING zz PRUNING FUNCTION BLANKS TABS STRING zz FF LEARNING FUNCTION BLANKS TABS STRING zz UPDATE FUNCTION BLANKS TABS STRING COMMENT unit section COMMENT default section COMMENT site section COMMENT type section COMMENT subnet section COMMENT conn section COMMENT layer section COMMENT trans section COMMENT time delay section COMMENT DEFAULT SECTION TITLE CUT COMMENT WHITESPACE default block default header SEVEN COLUMN LINE EOL COMMENT default def SEVEN COLUMN LINE EOL nc ACT COL SEP BIAS COL SEP ST COL SEP SUBNET COL SEP LAYER COL SEP ACT FUNC COL SEP OUT FUNC CUT zz SFLOAT W COL SEP SFLOAT VV COL SEP STRING W COL SEP INTEGER W COL SEP INTEGER W COL SEP STRING W COL SEP STRING CUT 26 site definition section site section site block site header site def type definition section type section type block type header type def subnet definition section subnet section subnet block subnet header subnet def unit definition section unit section unit block unit header unit def zz SITE SECTION TITLE CUT COMMENT WHITESPACE site block site header TWO COLUMN LINE EOL COMMENT site def TWO COLUMN LINE EOL SITE NAME SITE FUNCTION CUT zz STRING YY COL SEP STRING CUT zz TYPE SECTION TITLE CUT COMMENT WHITESPACE type block type header FOUR COLUMN LINE EOL COMMENT type def FOUR COLUMN LINE EOL NAME COL SEP
20. MUM_ODIM NO_OF_CLASSES CLASS_REDISTRIB REMAPFUNCTION REMAP_PARAM A 4 2 Grammar pattern file header i head o head vi head vo head cl head rm head actual dim actual dim rest cl distrib rm params paramlis pattern list pattern UPC PN n anything up to EOL FREE n mn J o 9 Vv INT INT INT INT EXP I INT INT EXP INT INT EXP Ee INT SNNS pattern definition file generated at FREE n No of patterns WHITE No of input units WHITE No of output units WHITE No of variable input dimensions WHITE No of variable output dimensions WHITE version number Maximum input dimensions WHITE Maximum output dimensions WHITE No of classes WHITE Class redistribution WHITE Remap function WHITE Remap parameters WHITE header pattern_list VERSION HEADER V NUMBER GENERATED AT NO OF PATTERN NUMBER i head o head vi head vo head cl head rm head NO OF INPUT NUMBER NO OF OUTPUT NUMBER NO OF VAR IDIM NUMBER MAXIMUM IDIM actual dim NO OF VAR ODIM NUMBER MAXIMUM ODIM actual dim NO OF CLASSES NUMBER cl distrib REMAPFUNCTION NAME rm params L BRACKET actual dim rest R BRACKET L_BRACKET R_BRACKET NUMBER actual dim rest NUMBER nc CLASS REDISTRIB
21. NE EOL COMMENT connection def THREE COLUMN LINE EOL TARGET COL SEP SITE COL SEP SOURCE WEIGHT CUT INTEGER W COL SEP COL SEP STRING W COL SEP INTEGER WHITESPACE COLON WHITESPACE SFLOAT COMMA INTEGER WHITESPACE COLON WHITESPACE SFLOAT CUT nc LAYER SECTION TITLE CUT COMMENT WHITESPACE layer block layer header TWO COLUMN LINE EOL COMMENT layer def TWO COLUMN LINE EOL LAYER COL SEP UNIT NO CUT zz INTEGER VV COL SEP INTEGER COMMENT INTEGER CUT TRANSLATION SECTION TITLE CUT COMMENT WHITESPACE translation block translation header THREE COLUMN LINE EOL COMMENT translation def THREE COLUMN LINE EOL DELTA X COL SEP DELTA Y COL SEPZ CUT zz INTEGER W COL SEP INTEGER W COL SEP INTEGER zz TIME DELAY SECTION TITLE CUT COMMENT WHITESPACE td block td header SIX COLUMN LINE EOL COMMENT td def SIX COLUMN LINE EOL NO COL SEP LLN COL SEP LUN COL SEP TROFF COL SEP SOFF COL SEP CTYPE CUT zz INTEGER VV COL SEP INTEGER W COL SEP INTEGER W COL SEP INTEGER W COL SEP INTEGER W COL SEP INTEGER W COL SEP A 4 Grammar of the Pattern Files The typographic conventions used for the pattern file grammar are the same as for the network file grammar see section A 3 1 1 28 A 4 1 Terminal Symbols WHITE FREE COMMENT L_BRACKET R_BRACKET INT V_NUMBER NUMBER EXP VERSION_HEADER GENERATED_AT NO_OF_PATTERN NO_OF_INPUT NO_OF_OUTPUT NO_OF_VAR_IDIM NO_OF_VAR_ODIM MAXIMUM_IDIM MAXI
22. NS To begin the tour let s start JavaNNS as described in Installing type java jar JavaNNS jar or if using Windows click the JavaNNS bat file After starting the program its main window opens As we have started the program no parameters in the command line the window is empty containing only the usual menu bar Also no network files have been loaded 4 2 Loading Files Use File Open menu to open an example file navigate to the examples directory and open the files xor untrained net and xor pat a simple network and a corresponding pattern file 4 3 View Network The main window still remains empty so choose View Network to display the network You should see a new window appearing schematically showing a network consisting of 4 units neurons and links between them in its main part Neurons and links have different colors representing different values of unit activations and link weights The colored bar on the left edge of the window shows which color corresponds to which value and can be used as reminder The colors and appearance in general can be adjusted through View Display Set tings which corresponds to the Display Setup window in SNNS 4 4 Training Network Let us now train the network reprogram its vveights so that it gives the desired output when presented an input pattern For that purpose open the Control Panel in the Tools menu The Control Panel is as in the SNNS the most important window in the s
23. T at least n Cea NESE CORNE LATE at least a blank Xt or n MANU E oe IS ee IS ee i H KRE l K EE NE pl EE ali ES nm magn PEN MR ON NE PS EE II PEN ap LL LL mpm ng LE Ion CSTRING a Io l V1 4 3D V2 1 V3 0 version of SNNS SNNS network definition file output file header generated at network name source files no of unites n no of connections no of unit types no of site types learning function s pruning function subordinate learning function update function 24 UNIT SECTION TITLE DEFAULT SECTION TITLE SITE SECTION TITLE TYPE SECTION TITLE CONNECTION SECTION TITLE LAYER SECTION TITLE SUBNET SECTION TITLE TRANSLATION SECTION TITLE TIME DELAY SECTION TITLE column titles of the different tables NO TYPE NAME UNIT NAME ACT BIAS ST POSITION SUBNET LAYER ACT FUNC OUT FUNC SITES SITE NAME SITE FUNCTION NAME TARGET SITE SOURCE WEIGHT UNIT NO DELTA X DELTA Y Z LLN LUN TROFF SOFF CTYPE INTEGER SFLOAT STRING unit definition section unit default section site definition section type definition section connection definition section layer definition section subnet definition section 3D translation section time delay section no type name unit name act bias gt position subnet Jayer a
24. UNIVERSITY OF TUBINGEN WILHELM SCHICKARD INSTITUTE FOR COMPUTER SCIENCE Department of Computer Architecture JavaNNS Java Neural Network Simulator User Manual Version 1 1 Igor Fischer Fabian Hennecke Christian Bannes Andreas Zell Contents Le NOU OH aset tetor O 4 11 FLOW LO read A UA lan pp ue duds 4 2 Licensing and Acknowledgements 2 2 2 aeree este pr dt eene ein ease dat 5 2 1 LICENSE Agreement dida 5 2 25 CNO IES ES roue e i rte 6 3 INA OS 8 3A RUME A ANN pk 8 4 A Quick Tour of JavaNNS een 9 4 1 Starting JavaNNS suit 9 4 2 Loading Plenos 9 4 3 View NetWork M M 9 Z4 Tranne NetWork o d kini Uo A Sn d mcs 9 ANA NODWODK rede qoo tole a entes e de ne ne 10 Zo Credane NO 10 4 1 Graphical Network Display ie deer e t tint Sn RERO RERO RINT 10 4 8 Training and Validation Pattern Sets 11 5 Network Creation and Editing uos tee perte eoe ter id sels s0 12 5 1 Network View and Display Settings eese 12 3 25 Tools for Creatine Networks se one a ul gelbe eta dou diucius 13 2 9 ECHECS e dte E ce A etd Ee 14 6 Pattern Management ss italia nda dei 15 7 Training and Pruning Networks ee ide ii tete 16 8 Analyzing O E 18 Bob EEO COM Panel oF ote ec Aa a gt le a SUR eme 18 8 27 Weights Paid E ie nine 18 Be ANaly Ze Pj ER ee a A R de 19 9 Loading Saving and Printing A oo diss 20 Appendix A Kernel File Interface 54e 21 A 1 The ASCII Network Pile Format ns
25. ar c a aee JE i FETES Se finition section ypeName unitName act bias st position act func ET ere i Me A e hae AS eae nl 1 00000 0 00000 i 3 5 0 ILI 25 2 1 00000 0 00000 i 9 5 0 ILI hidden 0 04728 3 08885 h 6 3 0 EE result 0 10377 2 54932 o 6 0 0 ILI ion definition section source weight 1 4 92521 2 4 83963 1 4 67122 2 4 53903 3211511523 33
26. bute JavaNNS as permitted above be liable to you for damages including any general special incidental or consequential damages arising out of the use or inability to use JavaNNS including but not limited to loss of data or data being rendered inaccurate or losses sustained by you or third parties or a failure of JavaNNS to operate with any other programs even if such holder or other party has been advised of the possibility of such damages 2 2 Acknowledgments JavaNNS is a joint effort of a large number of people computer science students research assistants as well as faculty members at the Institute for Parallel and Distributed High Perfor mance Systems IPVR at University of Stuttgart the Wilhelm Schickard Institute of Com puter Science at the University of T bingen and the European Particle Research Lab CERN in Geneva The project to develop an efficient and portable neural network simulator which later became SNNS was lead since 1989 by Dr Andreas Zell who designed SNNS and acted as advisor for more than two dozen independent research and Master s thesis projects that made up SNNS JavaNNS and some of its applications Over time the source grew to a total size of now 5MB in 160 000 lines of code Research began under the supervision of Prof Dr Andreas Reuter and Prof Dr Paul Levi We are grateful for their support and for providing us with the neces sary computer and network equipment The following persons were directly
27. ccur several times e X means that x can be omitted x n means that x has to occur exactly n times e xly means that either x or y has to occur e and bind strongest is second binds weakest e Groups or classes of characters are treated like a single character with respect to priority A 3 1 2 Definition of the Grammar The Grammar defining the interface is listed in a special form of EBNF e Parts between square brackets are facultative e separates alternatives like with terminal symbols x means that x may occur zero or more times e CSTRING is everything that is recognized as string by the C programming language A 3 2 Terminal Symbols WHITESPACE BLANKS TABS VV COL SEP COL SEP COMMA EOL CUT COLON separation lines for different tables 7 TWO COLUMN LINE THREE COLUMN LINE FOUR COLUMN LINE SIX COLUMN LINE SEVEN COLUMN LINE TEN COLUMN LINE COMMENT VERSION SNNS eleven different headers GENERATED AT NETWORK NAME SOURCE FILES NO OF UNITES NO OF CONNECTIONS NO OF UNIT TYPES NO OF SITE TYPES LEARNING FUNCTION PRUNING FUNCTION FF LEARNING FUNCTION UPDATE FUNCTION titles of the different sections 23 ne whitespaces PES only blanks or tabs C ln IC ENE N N at least one blank and the column separation U Inn le UP I NY column separation U Ia Ie Ust N at least the comma U In t No PA
28. character recognition with SNNS ART models ART1 ART2 ARTMAP and modification of the BigNet tool documentation about the SNNS project learning procedure Backpercolation 1 ANSI C translation of SNNS ANSI C translation of SNNS and source code maintenance Implementation of distributed kernel for workstation clusters Jordan and Elman networks implementation of the network analyzer Network pruning algorithms Redesign of C code generator snns2c Design and implementation of batchman Implementation of TACOMA and some modifications of Cas cade Correlation Java user interface design and development Java user interface development Java user interface support JavaNNS kernel port for Mac OS X There are a number of important external contributions by Martin Reczko Martin Riedmiller Mark Seemann Marcus Ritt Jamie DeCoster Jochen Biedermann Joachim Danz Christian Wehrfritz Randolf Werner Michael Berthold and Bruno Orsier 3 Installation To be able to use JavaNNS you have to have Java Runtime Environment or JDK which con tains it installed Java NNS has been tested to work with Java 1 3 Java 1 2 2 might also work though problems with file management have been reported in certain environments JavaNNS for Windows platforms is distributed as the zip file JavaNNS Win zip and as gzipped tar archive like JavaNNS LinuxIntel tar gz and JavaNNS Solaris tar gz for other operating systems Unzip unpack the file
29. ct func out func sites site name site function name target site source weight unitNo delta x delta y nz LLN LUN Troff Soff Ctype 0 9 I e 1 0 nos 0 9 5 signed float A Z a z integer I string A 3 3 Grammar out_file file_header parts of the file header h_snns h_generated_at h_network_name h_source_files h no of unites h no of connections h no of unit types h no of site types h learning function h pruning function h ff learning function h update function sections unit default section default section default block default header default def 25 file header sections WHITESPACE COMMENT h snns EOL COMMENT h generated at EOL COMMENT h network name EOL COMMENT h source files EOL COMMENT h no of unites EOL COMMENT h no of connections EOL COMMENT h no of unit types EOL COMMENT h no of site types EOL COMMENT h learning function EOL COMMENT h update function EOL COMMENT h pruning function EOL COMMENT ff learning function EOL SNNS BLANKS TABS VERSION zz GENERATED AT BLANKS TABS CSTRING NETWORK NAME BLANKS TABS STRING zz SOURCE FILES BLANKS TABS COLON BLANKS TABS CSTRING zz NO OF UNITES BLANKS TABS INTEGER zz NO OF CONNECTIONS BLANKS TABS INTEGER zz NO OF UNIT TYPES BLANKS TABS INTEGER zz NO OF SITE TYPES BLANKS TABS INTEGER zz LEARNING FUNCTION BLAN
30. dually For that purpose the user has to choose from the Edit menu which property he or she wants to edit The labels displaying the property turn to entry fields which can now be edited The fields are selected by the mouse and can be traversed by pressing the Tab key Pressing Enter accepts changes and turns the fields back to labels Figure 6 Create Links tool Changing activation values of units is useful if patterns are created manually 15 6 Pattern Management Like in SNNS patterns are organized in pattern sets which are stored as text files They can be loaded using the Open option and saved using Save data not Save E from the File menu Further manipulation is Mio E primarily performed from the Control Panel accessible from the Tools menu in the Patterns tab Some simple manipulations Trainingset bog varane HX S control Panel Lx Initializing Updating Learning Pruning Patterns Subpatterns Remapping function IERTEETRSDI IES M p Bx adding modifying deleting can be also validation set fbag varane mH 5S 14 4 DDI performed from the Patterns menu in the main menu bar Figure 7 Control Panel Patterns In the Control Panel a pattern remapping function and its parameters can be selected The two combo boxes Training set and Validation set are used for selecting the active train ing and validation set respectively Also when new pattern sets are created by pres
31. eneral modification Candidates output init Learn Maximum output unit error o2 I Print covariance and error Cache the unit activation v Prune new hidden unit Schwarz Bayesian y Minimize Figure 11 Cascade Correlation and TACOMA General cascade Correlation 8 TACOMA BE General Modification Candidates output init Learn Learning function IS ie rg y Parameters mpm mp TS XA wh Cascades fi 0 Learn Figure 12 Cascade Correlation and TACOMA Learning TACOMA can be set as a modification under the corresponding tab 18 8 Analyzing Networks Weights and Projection panels accessible through the View menu and Analyzer accessi ble from the Tools menu can be used for get ting insight into a network All the panels correspond to their SNNS counterparts and differ only in appearance 8 1 Projection Panel The Projection panel shows activation of a hid den or output unit as a function of activations of two input units The panel can only be opened when the three units are selected in a network view The activations of the input units are drawn on the x and y axis while cor responding activations of the output unit are represented by different pixel colors For the Projection to noName 18 iol x 49 Ke B 4 2 2 4 5 naName 1 a o E E Z o E v Figure 13 Projection panel chroma coding the same values as for the
32. etwork Simulator which is Copyright c 1990 95 SNNS Group Institute for Parallel and Distributed High Performance Systems IPVR University of Stuttgart Breitwiesenstrasse 20 22 70565 Stuttgart Germany Currently JavaNNS is distributed by the University of T bingen and only as a binary Although it is distributed free of charge please note that it is NOT PUBLIC DOMAIN The JavaNNS License is gives you the freedom to give avvay verbatim copies of the JavaNNS distribution which include the license We do not allow modified copies of our software or software derived from it to be distributed You may however distribute your modifications as separate files along with our unmodified JavaNNS software We encourage users to send changes and improvements which would benefit many other users to us so that all users may receive these improvements in a later version The restriction not to distribute modified copies is also useful to prevent bug reports from someone else s modifications For our protection we want to make certain that everyone understands that there is NO VVAR RANTY OF ANY KIND for the JavaNNS software 2 1 License Agreement 1 This License Agreement applies to the JavaNNS program and all accompanying programs and files that are distributed with a notice placed by the copyright holder saying it may be distributed under the terms of the JavaNNS License JavaNNS below refers to any such program or work and a work based
33. h The compiler determines the length of each row containing strings maximum string length t 2 Within the columns the strings are stored left adjusted Strings may not contain blanks but all special characters except The first character of a string has to be a letter Integers may have an arbitrary number of digits Cell numbers arealways positive and not zero Position coordinates may be positive or negative The compiler determines the length of each row containing integers maximum digit number 2 Within the columns the numbers are stored right adjusted Floats are always stored in fixed length with the format Vx yyyyy where V is the sign or blank x is O or 1 and y is the rational part 5 digits behind the decimal point Rows containing floats are therefore always 10 characters long 8 1 blank on each side If a row contains several sites in the type or unit definition section they are wri ten below each other They are separated in the following way Directly after the first site fo lows a comma 22 and a newline n The next line starts with an arbitrary number of blanks or tabs in front of the next site The source of a connection is described by a pair the cell number and the strength of the con nection It alvvays has the format nnn Vx yyyyy vvith the follovving meaning e nnn Number of the source e Vx yyyyy Strength of the connection as a float value format as discribed above The compiler determines
34. imulator because almost all modifications and manipulations of the network are done through it We shall also open the Error Graph window in order to watch the training progress Finally to receive some textual and numerical information vve can open the Log vvindovv Both are accessible through the View menu A sample screen shot with the windows open is shown in Figure 1 The Control Panel is contrary to the one in SNNS divided into six tabs each containing con trols for specific purpose For this introduction let us switch directly to the learning tab Here the user can choose the learning function set its parameters number of learning cycles and update steps and finally perform network initialization and learning The classic Backpropaga tion equals Std_Backpropagation in SNNS is the default learning function As you can see for each learning function default parameters are provided Learning is performed by pressing one of the buttons Learn current which performs training with the currently selected pattern and Learn all which trains the network with all patterns from the pattern set During learning the error graph displays the error curve the type of error to be drawn is set on the left edge of the window The error is also written into the log window 10 ES JavaNNS Lloji File Edit View Tools Pattern Window Help Exor_untrained lt 1 gt CIO EEE nl xl result A amp l 1 0 10 0 0
35. into a folder of your choice You should get 1 JavaNNS jar the Java archive file containing the JavaNNS user interface classes 2 examples folder with example networks patterns etc 3 manual folder containing this manual JavaNNS needs the kernel library in order to work properly If you run JavaNNS the first time on your machine a dialog will appear to ask you where to install the library JavaNNS then rememberes the location of the library by storing it in the file JavaNNS prop erties which is generated and placed into your home directory Personal in Windows termi nology If you which to change properties use Properties editor in the JavaNNS View menu You can delete the library or the JavaNNS properties file any time you want In that case the library installation dialog will appear the next time you start JavaNNS 3 1 Running JavaNNS That s all Now you can start JavaNNS by typing java jar JavaNNS jar from the command prompt or by clicking the JavaNNS jar file from the graphical user inter face 4 A Quick Tour of JavaNNS JavaNNS is a simulator for artificial neural networks i e computational models inspired by biological neural networks It enables you to use predefined networks or create your own to train and to analyze them If any of these terms is unknovyn to you please refer to a book about neural networks or to the SNNS User Manual this manual describes only the usage of JavaNNS 4 1 Starting JavaN
36. k view work view serves as a quick reminder for color to value correspondence Figure 2 Units are placed along an invisible grid in the network view Optionally above and below each unit diverse unit properties can be displayed Which ones as well as grid size chroma coding for units and links and some more data are set in the Display Settings panel accessible from the View menu This panel corresponds to the Display Setup panel in SNNS Display settings of view 25552 lt 2 gt Lx General units amp Links Grid size fo Subnet p Max value color mi z Value Min value color mi Text color m Zero value color Background color Selection color O OK Preview Default Cancel Figure 3 Display Settings General Display Settings comprise of two tabs General and Units amp Links or SOM for Kohonen tool In tab General grid size in pixels subnet number and chroma codes for different values can be set In Units amp Links the user can set which properties like name unit activation etc are to be shown above Top label and below Base label of each unit Also the user can decide if the connections are to be shown if their weights are to be displayed numeri cally and if the direction arrows should be drawn Sliders Maximum expected value and Maximum expected weight control the chroma coding for units and links since they determine which value corresponds to the full color as set in
37. low soon 1 1 How to read this manual Because of large similarities between SNNS and JavaNNS this manual covers only the differ ences between the two It should be read as a companion to the SNNS User Manual available from the WSI web site http www ra informatik uni tuebingen de SNNS We suggest that you first read the SNNS Manual in order to become acquainted with the the ory of neural networks the way they are implemented in the SNNS kernel and to get a basic idea of the SNNS graphical user interface If you are already familiar with SNNS you can skip this step and start directly with this manual In the next chapter you will find the license agreement Please read it carefully and make sure that it is acceptable for you before installing and using JavaNNS The installation process dif fers slightly for Windows and Unix machines and is described separately for each case After installing we suggest that you follow our quick tour through the simulator to get the first impression of how it is organized and used The rest of the manual covers in more detail creat ing manipulation and analyzing neural networks You can skim it in the first reading and use it later as a reference 2 Licensing and Acknowledgements JavaNNS is Copyright c 1996 2001 JavaNNS Group Wilhelm Schickard Institute for Com puter Science WSI University of T bingen Sand 1 72076 T bingen Germany It uses the kernel of SNNS Stuttgart Neural N
38. racteristics e the list of connections e alist of subnet numbers alist of layer numbers All parts except the header and the enumeration of the cells may be omitted Each part may also be empty It then consists only of the part title the header line and the boarder marks e g Eze Entries in the site definition section do not contain any empty columns The only empty col umn in the type definition section may be the sites column in which case the cells of this type do not have sites Entries in the unit definition section have at least the columns no cell number and position filled The entries rows are sorted by increasing cell number If column fypeName is filled the columns act func out func and sites remain empty Entries in the connection definition section have all columns filled The respective cell does not have a site if the column site is empty The entries are sorted by increasing number of the target cell column target Each entry may have multiple entries in the column sources In this case the entries number of the source cell and the connection strength are separated by a comma and a blank or by a comma and a newline see example in the Appendix B The file may contain comment lines Each line beginning with is skipped by the SNNS ker nel A 2 Form of the Network File Eries Columns are separated by the string without A row never exceeds 250 characters Strings may have arbitrary lengt
39. saved through Save data by choosing the appro priate file type in the combo box at the bottom of the dialog window For result files additional options start and end pattern inclusion of input and output files and file cre ation mode like in SNNS can be set Look in examples y amp taj al Es EH Start pattern fo End pattern fa Include input files Include output files create C append File name Jnyresutt Save Files of type Resutt iles res y Cancel Figure 16 File Save dialog When choosing files for loading in the file dialog window it is possible to select multiple files even of different types That way the user can load a network configuration and multiple pat tern files in only one step This currently doesn t work in the Linux implementation Print always refers to the currently active window Therefore anything that can be displayed in JavaNNS can also be printed by making the desired window active i e clicking it with the mouse and choosing Print from the File menu 21 Appendix A Kernel File Interface A 1 The ASCII Network File Format The ASCII representation of a network consists of the following parts e a header which contains information about the net e the definition of the teaching function e the definition of the sites e the definition of cell types e the definition of the default values for cells e the enumeration of the cells with all their cha
40. sing the second button next to each of the combo boxes the corresponding combo box becomes edit able so that the new pattern set can be given a name The other button adjacent to the combo box deletes the current pattern set from the memory Near the right edge of the panel in the pre last row three more buttons serve for modifying the current pattern set Their function from left to right is add copy and delete pattern Add creates a new pattern from current input and output unit activations and adds it to the current pattern set Copy creates a new pattern which is a verbatim copy of the currently selected one and adds it to the pattern set Finally the delete button deletes the currently selected pattern The current pattern is identified by its ordinal number in the pattern set This number is dis played in a text field between arrow buttons in the bottom right corner of the panel The arrow buttons provide for navigating through the patterns in the currently selected set Some patterns can contain subpatterns of variable length In that case the tab Subpatterns is enabled and provides for defining size and shape of subpatterns as well as for navigating through them This corresponds to the Subpattern window in SNNS Propagating patterns through the network is done in the Update tab of the Control Panel Same navigational controls are provided as in the Patterns tab Besides the button between the arrows propagates the current
41. sis 21 AA Form of the Network A e ete sen 21 A 3 Grammar of the Network e e a tte 22 Ad Grammar of the Pattern Eales ti Gas dhen ee te riada 27 Appendix B Example Network File ann ic eee tnt oet e en etn d si 30 BL Bxample A nn A k he 30 B 2 Example nn nn nt nn eae Ant nt der a a les tan an 32 1 Introduction Java Neural Network Simulator JavaNNS is a simulator for neural networks developed at the Wilhelm Schickard Institute for Computer Science WSI in T bingen Germany It is based on the Stuttgart Neural Network Simulator SNNS 4 2 kernel with a new graphical user interface written in Java set on top of it As a consequence the capabilities of JavaNNS are mostly equal to the capabilities of the SNNS whereas the user interface has been newly designed and so we hope become easier and more intuitive to use Some complex but not very often used features of the SNNS e g three dimensional display of neural networks have been left out or postponed for a later version whereas some new like the log panel have been introduced Besides the new user interface a big advantage of JavaNNS is its increased platform indepen dence Whereas SNNS was developed with primarily Unix workstations in mind JavaNNS also runs on PCs provided that the Java Runtime Environment is installed As of writing of this manual JavaNNS has been tested on e Windows NT e Windows 2000 RedHat Linux 6 1 e Solaris 7 e MacOS X with more to fol
42. the width of the column nnn by the highest cell number present The cell numbers are written into the column right adjusted according to the rules for integers The column Vx yyyyy has fixed width Several source pairs in an entry to the connection defi nition section are separated by a comma and a blank If the list of source pairs exceeds the length of one line the line has to be parted after the following rule e Separation is always between pairs never within them e The comma between pairs is always directly behind the last pair i e remains in the old line After a newline M an arbitrary number of blanks or tabs may precede the next pair A 3 Grammar of the Network Files A 3 1 Conventions A 3 1 1 Lexical Elements of the Grammar The lexical elements of the grammar which defines network files are listed as regular expre sions The first column lists the name of the symbol the second the regular expresion defining 1t The third column may contain comments e All terminals characters are put between e Elements of sets are put between square brackets Within the brackets the characters rep resent themselves even without and defines a range of values The class of digits is defined e g as 0 9 e Characters can be combined into groups with parenteses e x means that the character or group x can occur zero or more times e x means that the character or group x must occur at least once but may o
43. to turn the text fields to labels again 4 8 Training and Validation Pattern Sets To see how two pattern sets can be used for training and validation load two pattern sets from examples directory tramMAP pat and validMAP pat In the Control Panel tab Patterns select trainMap as the training set and validMAP as the validation set Switch back to the Learning tab and train the network During training two curves are displayed in the Error Graph one who s color depends on the number of already displayed curves and which repre sents the error of the training set and the other pink one which represents the error of the val idation set The validation set is normally used to avoid overtraining of a network For more information refer to the SNNS User Manual and other neural networks literature 12 5 Network Creation and Editing 5 1 Network View and Display Settings Although not necessary it is recommended that a network view is open when creating and editing networks Network view is opened through View Network menu The network view displays a visual repre sentation of the network which comprises of units and connec tions links between them Units are drawn as colored squares with 16 pixels side length and connections as col ored lines For both units and links the color represents a value activation for units and weight for links The colored bar on the left edge of the net Figure 2 Networ
44. une input units Minimum error to train pi Prune hidden units Initial matrix value o 000001 Prune Figure 10 Control Panel Pruning ing is performed by pressing the Prune button Cascade correlation and TACOMA learning are the only exceptions to the concept of the Control Panel being the central place for manipulating networks Because of the large number of parameters needed by the two learning methods a separate window accessi ble from the Tools menu is used Contrary to SNNS vvhere parameters for cascade correlation and TACOMA are dispersed between the Control Panel and the Cascade vvindovv in JavaNNS the Cascade vvindovv alone covers all data and parameters needed for apply ing the tyvo learning algorithms The window is divided into six tabs Tabs General Modification and Learn cover the parameters set in SNNS in the section General Parameters for Cascade of the Cascade window Under the Lear tab of the JavaNNS Cascade window the learning function together with its parameters and the maximal number of cascades hidden units generated during learning are set The Init tab is introduced for convenience and provides for initializing network Tabs Candidates and Output correspond to Parameters for Candidate Units and Parameters for Output Units sections in the SNNS window The default learning method invoked from the window is cascade correlation 17 E cascade Correlation 8 TACOMA MEI G
Download Pdf Manuals
Related Search