rit doc, 507i aft ., tor unlimited. v f o o · cli unclassifiedounlimited 0i same as rpt cj dtic...
Post on 14-Oct-2020
1 Views
Preview:
TRANSCRIPT
RIT DOC, AD-A277 507I,a. Ri I¼ ;?I Y CI SSI r Ir- C
Unclassified .
2a. SECURITY CLASSIFICATION AUTHORITY _Jw -7
aft________ _., krpL.uvea tor puDLic release; distribution2b DECLASSIFICArION /DOWNGRADI ED C 019 unlimited.
4 PERFORMING ORGANIZATION REP M8IER(S) S MONITORING ORGANIZATION FREPORT NUMBER(S)V F 5 #69 o o6a. NAME OF PERFORMING ORGANIZATION 6b OFFICE SYMBOL ila NAME OF MONITORING ORGANIZATION
Institute for Brain and (if appiicabfe)'e - Personnel and Training Research Programs
Neural Systems Office of Naval Research (Code 1142PT)
6c. A'DORESS City, State. and Of-Code) 7b ADDRESS (City, Sadce. and ZIP Cce)
Brown University 800 11orth Quincy Street
Providence, Rhode Island 02912 Arl-iagton, VA 22217-5000
8a. NAME OF FUNDING ISPCNSORING 8b. OFFICE SYMBOL 9. PROCUREMENT INSTRUMENT IDENTIFICATION NUMBERORGANIZATION (if applicable)
N00014-91-J-1316
8c. ADDRESS (City, State, and ZIP Code) 10 SOURCE OF FUNDING NUMBERS
PROGRAM PROJECT TASK WORK UNITELEMENT NO NO NO ACCESS:ON NO
1 1 TITLE (Include Security Classification)
Proceedings of the NIPS'93 Postconference Workshop on Methods for Combining
N Ij~1 M~fitjnrkr'z12 PERSONAL AUTHOR(S)
-- ij-.ael P. Perrone
13a. IYPE OF REPORT [713b. TIME COVERED D14 DATE OF REPORT (Year, Mont. Day) 15. PAGE COUNTFRTechnical Report ROM ro March 15, 1994 51(2-sided)
16. SUPPLEMENTARY NOTATION
17. COSATI CODES 18. SUBJECT TERMS (Continue on reverse if necessary and identify by block number)
FIELD GROUP SUB-GROUP AveragingEnsemble Methods, Committees
05 n
19. ABSTRACT (Continue on reverse if necessary and identify by block number)
TV',is report is the proceedings from the NIPS'93 postconference workshop entitled "Pulling
It '.ll Together: Methods for Combining Neural Networks." The goal of the workshop was to
examine theoretical aspects and heuristic methods for combining existing neural network
algorithms to generate systems which have improved performance properties.
This report includes an outline and summary of the workshop as well as the slides used in
each participant's presentation.
D2TIC Q UAL;'r' • .
20. OISTRIBUTION/AVAILABILITY OF ABSTRACT 21 ABSTRACT 'ECURITY CLASSIFICATION
CLI UNCLASSIFIEDOUNLIMITED 0i SAME AS RPT Cj DTIC USERS Unclassified
22a NAME OF RESPONSIBLE INDIVIDUAL 22b TELEPHONE (Include Area Code) I 22c OFFICE SYMBOL
Dr. Joel Davis l (703) 696-4744
DO FORM 1473,84 MAR 83 APR editon may be used until •xhaulted _ ý kjPiTY (1ASýSI i( AI()j()r4 1OHf ,(
All othor ed•tiOns are obh.olete
94 3 29 024
DISCLAIMII NOTYcl_0ý
\ M\
THIS DOCUMENT IS BEST
QUALITY AVAILABLE. THE COPY
FURNISHED TO DTIC CONTAINED
A SIGNIFICANT NUMBER OF
PAGES WHICH DO NOT
REPRODUCE LEGIBLY.
*~~~~Z j' .Z, m ~ m- -mt -
itt
-E-0E
-x- Fi
f 7c 0~ ~
r, 7t zi 6 2
T" C~ m
c 7 it *
72 it4
-. ~~ - . ii t r C
C~ ~ ~ c-- -
-0
~Z~;-0
o~~~~~~ -ctt-.zizCi
*'C7ý &
C-1
-7. -*71N
U 70
-o-
-(.Oc
I - -1' %r
0? or a?
* -s
fn -
cU5
z- -C-- -C
~ ~o -c C
~ ~1' U ~ z~En
E+
Et
fn 0 0
- - e~ -~ (
-S O~ ~ ~ =
bo~
&~~~~~ E~ i~- -
-~E
rc.l
-CC
CLj
- .0 0E0 C
0 0 0 -= 0 0 0
-ce
oo
0 >
00
'40- .
'~ G. -~ 0 LIc
-~ ~0H ocn
00
C OCT
0) -ho'UC
o 2 2 -
U -0 -, QZ rn ~- t
Ct2fd Z ~ .3
ol
CCl)
v I v I_Iv 1-j-
rj
Z: C L
+ c
oV
+
L~z L4
0
0
w H
.0
- H HH H
H IEL.
0 � __ z�II II
III
s..i HU)
bO '� x
o -�
*G3-�
2 -� -�
H I.�
IIVJ�i -, -
0.
+ .�z-H -�
H
IIH -� -v
'-III L�N1i Iii
U) I��) �-I-n, �
- ��
ii � E-o
Q.� - 0.- C� 0.0. -0� G.)
0
* . I
_- A I- III A I V
S • Al ± Al AIIlV
o ¢- lL - E •
S- - II
* =
E -lV- V'--
I-" C•"' • "-" '
t-, ,
P- 00C
Q0
L
> U50
Cd 'CI C -
-1-0
.~ ~- ce - CN
H 0 0)
LO)) 0)
o bi) v
ce (U
(CLI -oz Frix = *
> T. C
41 4z
.- ,
/i3S
/a*
3SW
000
V¶ V
1.2.2ZT0
w as
4.) 5 C7
cn4
.61=
-owD
.00
moii
0.
a U
C9I
404
"4 .• 0..,
0 -. 6
z90'a k -
E-4-
"Z ........ ...... .........
z... .
0W0
C) * 0
~ 4 0
ctd
xr
Q) C
WD E10 0..
W 0
-B U
~ .~ E
S rJ-~ - ~ S...0
0 z - b
wr -0 1-41-S... ... _
to
tt >~ClC)
0-
0 i -4J ~v4
IL"- 411vVI X1 dj w
C-C) (V -5 I
U ZJ-4J
CM\ >( 4) 0
C ,-.- 6T
II II
5,) -51
QJ d' -'H
('7 014H' •~
-4 J- 73
4Jsa LL (
0 Io Ot .
SR ~ L-3 ~C4-00 wj~ YL
-*ý 0)72 a
a-, (I, Cq-
'14
C4-1 (4ý
7; X
4i (IL'
)< I
4-J
4 r LA Y
GJ4
441
-,U
-41- -j
I- - 0~ -4
ci5- 4U _ o
ý-v yi 4-Jn~
_ 4 J -4 J
y - _-z ýK (I
p•LL
>I
0 o -
-S-5
C-I T4'*- y- ?7 w -4,
G- rI, ~ ~ .
1-4 ~ --- J
XI -
+2 it -
(-Ii
t~-i ,J
L4!, -4 41
-'I-li~~ LL, . -
+4J- I -
QJ~~~~ \1 ~~Q
-4J-
6D1 -\
-P 0
-~-41
L4 0j
- 0P
'-4)
4J _ 3- ~AJ -4J
,~ 241
QjQ
4J )<-)-
d I-
2 -i
CiL' < Cl.[i-- J >e, -
~ -~ H
I '
4j qj r
171-
Q) --
CO1 t 1-1 -
0jO4 WI
qi LL L
qj~ ~LLz-> qd j
LI4 6t)
S XI
LI +
cpQLI
-4Ji
-~ LI
r Il
I I-,5<W4ixi +
C',
L Jo
Ii V S)
NV~
C-+j
ItA
L) vrKj
~Q -4J~
it H
IV4-
'-J ~ i- 1-
xi
QX1~ +
1-'
1+1
Q.4-
-4 ii.i F~- -~ k1 ~ :-n
14U4--41 14.
U 'I-
0
xZ rnI
ILl
cs
-+j
-1 - -p
r're
-414 qj-~ .
4440
Icý I '-77-
4- x II -0
N -. J
I cJ( o
400
rlC'x-
14
14-
L/)
41P
Cqz
III tl
-t to -
oL io eo t l o • L 0o o o ooSOL 00 00 00 00 oo ) o• go *o .o •o oo (' )
0 0
0
Ir I
0l ,0 g.0 (0 0 0 l" 0oI 0go o' 000 01 SO 00 '0 00 0 0 Ot $00 gO ,00Z ,
goo
'-0
i Hi t • Ill n I • ug | II m | a it | i
II F
Ul)
..4 ... ....- -i t 0 10 t
6r ----- Ln to t0 1, 0 -0 00 C0
r r
ZL 0 to to to z' oo
goL il IL
iEll
0 . 0 0 G # 02~0' G O G O 0 ZO 0 0 O Z' S , 0 : ' S O ' 0 0 1
0, o 1 0 go .o zO oo0
I 4I
c~cJ
ii
ti I rI 3
SZ O i O ,O-0 O
0 0
0 0
IGI
0 1 6 0DO
1O00 0100~0 0
to 0 0
s0o 40 o01J to Do0
00 0
0 0 00 0 O 0
9' 9 0 O 0 000 0 0 0 00 0 01 0 go '0 Z0 00 1, t oo 9 0 . 0 0 0
T-ii
ii -t
4J4 > - -. \(-
0 1ý-4 b I
0 40.4~ 0
w) M (1 0 S
)-. e--4
La P4 0)14 .-6D 4-7 5 Q)
W s- -X 0
04 a)) w ..'.40 0..L
L)a Q)v *"
.- 0. s-4 (ncow-c! 01 -4 40(1
C.) 4. O c
W ) 0 Q tQ
4-
ce 0 0e(L
a)o
040
~4-5000 40
P4J
5--
* 4-
00
0 h91. -n
U, I- IC'
o '- ~ 4. -oT
uj 0~ -~ ~ ~WD
II b .0
0 ~ b U2 ~ rw 0
V -~ -- "
0
0 0
4-D .0 0 '-4
P4-
0-4
bOO
-4~
0 >. .a4-
ce 5- 0
00
(1) ct 0 -
a4S
U) -
0) 4P
9)4 0Ux 0.0
24ý44. $.-
0 .6.20 x *4I-6-D.M- 44-4
0 000
bO *4 0CV.U) 4-a 0_6.
-. 6..4 6.4C. )
0. 0 b0 . 04 - -0 x
0 bO0U) 0O 00 0.
0 .=
0 to0 4
0 4-'In)
r~
.46
CI.-
C)CQ0)
co 44
'U Q)cebO
1-
1-4~
"0 z-
>i U 1 4. .,-
cd Q00 40 1. Q)w >
I) Ll.I0A 0-4. w Cl)ý
bb0'
0 0 0 0 6l 0
-H -H
-L c i 06 0ý t
oz C14
-H -H - -H -H
I-
00 0
00
S--.
__ t-j 0
-c$ -0 10 0~ -c) _
ci. .4- 4 r
C)~ C) 4' r
C14.
I I I I I I I I I
C) mc
00* -
000ca. O Co - O~(
Coec
ce•0
"0~
0 ow 0
uCi
PIN
IJ
II
! I*t*J!W V°
Q))
00
- ..0. 4.4-4 ~5-~0 11~~ 'o
CLCl C C.)
)0 $-4CL,. s. C~ C) 4.-j
.0 P5 ' O
c~ - -4
Q) 0
Cd.
-0 Cd5-4 C.) I
oý c) -
C) . 0
f--o
-6.' -
4Q) .)
0 04.C.) h.oC5.4 -
>- >)
CNO -
oh
tko bo-.0
pooQi- a sIc
ch' .4.
3..~. .-. .-
ct Id
14-I4
ý-4 w~
+;
C-4
xI-
x;
:y I I
z ~ ~ ~ ~ ~ ~~I 'X C)-d. e(')S 213=70
Hwý -0 0xg u UI
Orr;
MniOU2ra (IN Sx'g),91N
-ho
bC,,
bj4
0--
>+
C,,
Cf.)
C,, ' -4
.4- -4-
(1) 0,
Ct Q 0
'-'e-
C) cr.)-
-M
IIe
U)
UU) cn j 0HUJ < wLLU U) n Zcl
z < E o
0-00
L LU W< <D 2: W 1 -0 (0C)> L
LL
C',L
0
I-i0 wwc jDCz : 0 o<
m~ cc Loow> -I) w cc -wo ( ( 2 Z 'Z M
LI 10I--0
z
C,)l
ZO
Z U
Z
cr..
LU U)
0
0- 000
0 wR 0W >
= LL
C/) cc
F 'C/) 0l mr0 w J
0
a0
CD MZ-=w E2C/)0 cc. N
0 1f () '- a.) Ec
oo 1-MC
E oa CL N EL- -. --Oa
C4OX a-.C E Cc z 146 I 0l cc 0CD LLw oa- Z 1 00 '- Cc$ 0 LL cc
0 L LCcC- U
o ooLLz0 - -
00)00)00
00)0a)00
zwj10z zC.) U)
N10
wj w
0.1
zo
oz zoo m LULC
LUU
00
LU N 0
0C- ~ C ~ Cz
erj c' -
LLI-
ui D 0z Lz
LUU
wn C/)
CN/
zo-
crn~vWNfl SN83llVd M3N
ONIISO0O9 IVNIVE1O
05 cf0
z 0'
0
z~~ 0 Lmw - 0) 0
0 -0
z c~~0
z
0wI .),
0n
uj LU I-
z 3 C/)
R 4c04'V- C1 m czLU 00 I
LU N
w
Z< 2r
z0
0
0
z
Z N
00
0
coo
LU00J C
c, C
0 0 < N 0CC a:C/) 0
0 I-i 20
ow C/:zLU]
FU 0Uc
0cl 8 W :v l C~ C-4 -r '-1-l
0'-
C,z
LI-
z ~0 c
LI-
w C/) m
z~ LL
w~ L)
uz0. I-
03NIUI~n SNH3I~d 3N-V
DNIISOS (33:lla)
0,zo :a . R, R ,
LL x 0 0
o. C/) -zp Jýr0 W'
Clo WL w iLo l Cl, F-. cnCl) 0D
0
zm
z z znN c LU
w CCIw wU
0 .L
LU 0z C) 0
,-, 0
_j - 1 0J LU 0SLU < LU NLU ~ N~0
ZU < ~0 LU
uj < 0
0 <
oCC
CCC
LU L
LU0 0 00 ~00
N 0 000 0-
-LU LU L
0) 0 N
< R 02
Zi._j 0D.0
D0 a zHL <
0 0 C a0
04cl co C : D . C%.J C>
0 0i 0 0;
R
0
Fwi 0~00
se 0 2 0~
0 0 u 0 F
0
I0 0.
cc DI
(N (N N -4 ,4 ,4 .4~
00
z 00
(0 <
0 0 L)<J (52 N
0 0u
z < E LLJ~z I--
0 ~ a.< U
0 E3 0* 0 Z1*~ z
;20a
0cF1- L
00
0 00 0
0 Cicnco th
zza:o0 cn(n 0
0 c0 C4 lo .
Lm( 0 .
wU N L wz) z 0f0Z Lii Z o
0I 0f IcC6 ccJ m\ 'N di
LU ccZ
0
LLJ
CV Z rC C/)H
0U 0W0
H -J ~C4 zl IW
0w Z >
o 0 Zu0 2 Wujcc l)- > z crzr
Ml Cf) 00z
CC L
A Commercial Application:Extracting Document Content
from Images
Christopher L. ScofieldHarry Chang
Ed Collins
IifINestor, Inc.
Structure of the ProblemOCR is not really a problem of characterrecognition. It is really language processingfrom images:
Character context drives segmentation:
Lexical context drives character interpretation:
iO Nestor, Inc.
Structure of the Problem
Lexical context -"',es character interpretation:
- clone?- done?
Application specific rules drive interpretation:
54Y293?5" d y 72k -3 54Y2A3?
J4Y2A3?
Ii Nestor, Inc.
Structure of the Problem
Document structure drives syntactic and lexical possibilities:
Company Name
Street number Street name
City State Zip
I 7 Nestor, Inc.
What are the possibleapproaches?
Single Network Architecture" [Keeler90]: Combined segmentation and recognition;
,, [Fontaine92]: RNN trained on pixel-columns
- Pluses:
, Makes no assumption about structure of problemAutomatically trains each part of problem
- Minuses:
Scaling problemLack of modularity: application dependent
If iNestor, Inc.
Possible methods
* Multiple Network Architecture" [Gouln92]: Neural network segmentation for map
processing;
, [Scofield92]: Context-driven segmentation,recognition
- Pluses:, Each module can be built in a minimal fashion
Only some parts need to be changed for newapplications
- Minuses:" Assumes prior knowledge, Must be assembled in a piecewise fashion
Credit assignment problem
IFJNestor, Inc.__
A Multiple Network Approach toOCR for Handwriting
Task decomposition:
S~~~ocn. .'Struture--
S.Reog~nition LSegmentation
I lNestor, Inc.
Neural Network Segmentation
• Neural network is used to assemble a treerepresentation of the image:
(1) Classify all blobs Into "Character", "Noise", and "Mixed"
(2) Recursively segment "Mixed" until decomposed intoonly terminal nodes "Character" and "Noise"
(3) Compose a list of possible alternative segmentationsFragmented charactersOptimal window adjustment
F Nestor, Inc. _
Segmentation: Step 1
e 3-layer BPN trained to classify blobs into:"Character" "Noise" "Mixed"
it* * ,o.
e Use connectivity analysis features [Hu62]including:
area, perimeter, number of holes,
area of holes, principle moments,
aspect ratio, etc.
ffl Nestor, Inc.
Segmentation: Step 2S"C",c, "N" are terminal,"M" parsed with quad-
tree analysis [SametO]
Re-classified at each _
step: •j_ J
f 7 Nestor, Inc.
Segmentation: Step 2 (contd):- Hierarchical Agglomerative Clustering [Duda72] groups
terminal nodes; re-classified to ensure still terminal:
Step 3:a List of segmentation alternatives compiled for
classification into characters- Multiple "cuts" of characters provided for later analysis:
IQiNestor, Inc.
Segmentation Accuracy
"* Test set consists of 5,654 HP/MP charactersin 1,236 words (46% HP) selected from 53real-world documents
"* HP data consists of live forms withconstrained HP, unconstrained HP, run-onHP and some cursive
* Character segmentation accuracy:(First choice correctly segmented)
Segmentation Network Correct Incorrect
"Blob" features (7-10-3) 92.7 7.3
I 7 Nestor, Inc.
Character Recognition:Overview
* Segmentation alternatives are processed forcharacter class
° Use two static feature sets: (cf - [LeCun90])"° Three-layer, feedforward BPNs are used as
estimators of a posteriori probabilities"° We have employed three types of hybrid
networks:- Glue networks [Waibel88]- Parallel Experts [Reilly87]- Hierarchical Filters [Reilly87]
\ffO Nestor, Inc.
Character Recognition:Feature Extraction
"* Segmentation alternatives are converted togrey-level: gaussian kernel estimated from linewidths
"* Pixel Feature Set:- Pixel map is sub-sampled with grid producing coarse
map (100-element grid)
"° Edge Feature Set:- Edge-map produced from grey-level gradient estimation
[Roberts65] (4 edge directions)
- Edge map is sub-sampled with grid producing coarseedge map (30-element grid)
If lNestor, Inc.
Character Recognition Classifiers° Features used to train 3-layer, feedforward BPNs
Data Set # Authors Digits Alpha U/LC
Train: NIST 1,3; Propr. 2600 265,000 120,000
Test: Propr. 4,767 12,932
In addition to using "Forced accuracy", can use aheuristic which models high cost of errors:
"Figure of Merit": FM = 100 - 10(%E) - %R
Numeric Network FM Correct Inc. Reject Forced
Edge features (120-32-10) 95.15 97.04 0.21 2.75 99.01Pixel features (100-45-10) 93.41 95.87 0.27 3.86 98.59
FflNestor, Inc.
Classifier Analysis
e Using "rule-of-thumb" e = W/T [Baum89]:
Training set: T - 265,000
Edge Net: W - 120*32+32*10 = 4,160Expected test error: e = 1.7%
Pixel Net: W = 100*45+45*10 = 4,950Expected test error: e = 1.9%
fi Nestor, Inc.
Character Recognition:Parallel Experts
"* How to combine the results from two networks?
"* Could vote if have many "experts'. If only two, thenaverage activation (probability) vectors:
Pi = 1/2(Pei + PP1) 0
Numeric Network FM Correct Inc. Reject ForcedEdge features (120-32-10) 95.15 97.04 0.21 2.75 99.01
Pixel features (100-45-10) 93.41 95.87 0.27 3.86 98.59
Parallel Nets: 97.25 98.20 0.10 1.70 99.39
IFiNestor, Inc.
Alphanumeric Character Recognition
"* Support full alphanumeric HP"* Natural decomposition into alpha and
numeric subnets
" Use glue-net architecture [Waibel88]:- Trained 3-layer nets for alpha (u/I case) and numeric
- Freeze middle-layer weights, route activations to output- Add-in new "glue" layer to resolve Inter-class ambiguity
- Train second layer of weights and all glue-cell weights
alpha numeric alpha numeric alphanumeric
[ iNestor, Inc.
Alphanumeric Accuracy* Results: Forced
Network Architecture FM Correct IncorrectNumeric sub-net (120-32-10) n/a 95.15 4.75Alpha (U/L) sub-net (120-120-26) n/a 91.41 8.59Single Glue Net(l) (120-210-36) 51.24 89.24 10.76
Hierarchical (Super) Glue Net 54.04 90.80 9.20Single Glue Net(2) (120-210-36) 54.46 89.98 10.02
SIumeric
Pixel features
Edge features
fIINestor, Inc. Eg~et
Glue Net Analysis
Digit set: 265,000Digit Net: W = 120*32+32*10 = 4,160Expected test error: e = 1.7%Alpha set: 120,000Alpha sub-net: W = 120*120 + 120*26 = 18,720Expected test error: e = 15.6%Full set: 385,000Glue weights: W = 120*58 + 210*36 = 14,520Expected test error: e = 3.8%
iONestor, Inc.
Character and Lexical Context
Word recognition: determine the best stringinterpretation given all sources of knowledge:
segmentation alternativescharacter recognition probabilitiescharacter transition probabilitieslexical context
el~o-n-e O --V 1 D7cl 0/m--P D H U"I \P
Character and Lexical Context
"* Use the Viterbi algorithm [Viterbi67, Forney73]to select the character string with maximum aposteriori probability
"* Let maximization of word probability driveproper segmentation [Bozinovic82]
"* Problem: VA can produce lexically incorrectstrings. Post-processing with a dictionary canproduce word which is not MAP.
"* Solution: Use lexical context to trim pathsfrom VA search ensuring that the final string isboth MAP and lexically correct (Srihari83].
IO Nestor, Inc.__ _
Application Context Processing" Some applications are alphanumeric but not
part of lexicons: A1800138- Inter-character statistics are specialized, hard to learn
without large set
"* User-definable syntax selects which subnet touse for each character position, trimssegmentation alternatives to match syntax
ZIPCode
(FiN es tor, Inc.__ __ _ __ __ _
Status and the Future
* This architecture is the basis for the productNestorReader
* To be supported on Nil 000 neural net chip
* Extensible to character-based cursiverecognition
- Now developing much larger training and test sets for run-on HP and cursive
- Developing stats. on character and segmentationaccuracy due to character and lexical context
IO Nestor, Inc.
L4--
"4- ~ ..zl
Is
44
I-'1
LLb
0041 0061 OCSlt oss 0
Wddl
out o t OGGt 096
I0
L~ddi
v 40la, CC1 Vd OI
see$ 006 OS 0L6 I ~SWL 096 L m 1 06
LA
(cc~7~'A ~ 'A~?~
IL
ID
0..
a - - CL0 I
I. U(
on
3ONYHO~~~ ~ ~ ~ ~ IN18d3NH N03 0Y)C3@
%WdI
,addA
aA IX
"I*-
Lu
a S
1.4 -0 -TO"a
-a-AR(14) hewraed/-a-Unew Rg.
0 1.2 pa Net
* 0.8
0 2 4 6 8 10 t2
Prediction Honizon (Months)
-cP .- ,".~ , ýy ve s t";
.~pb
-c dS9
3 'C ~x *-
.2"z z i ~ e i l l , Ict wa?5.1n;2 x:.9
'11 - . I 1 i>1________ r il
ii _______goo
dzIQ 3
y L4
Up x
'J. ~A
L44-
oc 9z 0.1 SLO
9,
II
Pq 4-A .
J9SA
99~ 99z10
~AM
JeeAS
9z
*~ ~0
9'0
jes Buiuieal jv9A L k
Ip -- 7yr training committee vs average
0.65 -
Drawsocomijt~tee -
0.6
0.45 / ,0.55
0.35
0.3 /
0. 25 t'
0.2 ""
0.15 'so 82 84 86 as 90
r. ÷u• a- n,,, v, C..ess.1
IP -"7yr train'ing :coxmmittee vs averag'e0.65, ,,
r ]aw-
0. 6 "commit tee". ..
0.55 '
0.50.45 / '
0.4 / ,
* S
0.35,'"
0. 3 ,'
0.250.' "
0.15 ,
80 82 84 86 88 90
•,,' • •; " _•A )-
•." ,-, . . - /, . "
G4..
,'
N NL
V -N
N Z -
74A.
'S4
zSj"S
-L
\lLk ,
1';;ý /4LA,
K.L
' ' \
~-\ ~ -3-11
- . ,I-.
, ) , ¢ .
1,,
S - - . • • • • sal '-i-a i i i• I I II Ii I ws
I% V
* -S 44 I
'.4-. �'
-V.. 4. t (1%�
0� '+,� 0
6) S., 44.4.. at c *
4 9%. I,
0 *�Ub C � 4-
'a 4OD
-� 4 '1.10
GD
k
cr�
9r0
4
'I'I
V
'4.
61
4
Fb.1;'4,'a6)
(A
C C 4�� E4-� -
�1% � SI
.4- -
SI SD 'Sfri � Esm. � e.3 4
4�. 6 ' � �m.4.. 4 SI
'a* SI � 4 � .,O 4*
'. 8. IS 4ID) * q� � I-,- *
'F *� C4�' � -I-
.% SD
E *! �.g1.�
S.. fri *.3 em..4 -.�.
'-.-.Z�- C
4- C �., 0
-m-m--E.-�--- - .m.� *,� �* IL1 � * '
"1 i- 34
*0-30
9%'C4 .� .�
I'.'. C 6 C -.3 6
.�* *% qL�. C �
'S. .* -� d1.. * 4 *1�
6 C
1d� jh d £
4%, .�.4
1 Sq34 �. %%6 6 \� I-
0 .� Ti �
N £ 6 6�7'- d j��4
V N -. 0 .. .-. .. 0.. � I ,-�0 Va 0 C
*4D %.. 'p *b
9 %%
ID Sb
0 0 0 0 * "I
"SI .
4 Ii, 4� ¶1V *0 * -
olu 0� 4, *'� �
U + �SD Q
£V .. 4
-44 SD,) �
( g% * bl
C �D *i' .¶�N . 4 .4
�mim. �4V ( UGD.,. 4
urn6 .- �.. .�
GD X9 1- �
� a *V% o-te %.
Qb 3 ti�. 4�*- 0
(Si
GD� -,-
'p .5
9'I
I
a
b �: �J1
( . - (Id aI Lb1'
S
*1'I
'. 'p.
'PD C.%
j C
� -
f) �-.
04
CCo
0-zU
Cu _ 0
Sal cu.s c0 Q 0 1 6
Qu c -ý 0 *ý$ -P
Poi1 No-
N0UZI
_~ Eu
g u
E CL
-0 M 0- 0
CCL)
o -2
0. 0
.6- -6-o .6.2
ga. bO 0:Id
.., 0 %0-4
- 0FE -4- >
0- =1I -- 6.
CIS , w ~ w.Q) bO 00 ~2~'.o xi,~.
00 'H 0e,_0
o am
0 I,
/ o1$-4 W Lo
CILI
o
$-4$-4
0)4.2 &4~;
b\ 0
06.
w w0 0bOe 0
0 0 0
bOO
Q ID
S CZ)
6- V
ce ~ .2 8$- go,-
a)o ;p ,. a
W~ Q)
.40v 1; it!
*I-
P4 -12*
Mi! V
fil
.6.
0 b
1.0
%- 40
-CC.
CGO)
0
(V 4tMl A. C
gy 0-t- ý
i tip
'/14( 4 * 3. . .. ... ..
3Q '16 1,- 0 vCf
.%b4
I-6Q
""A CL4~
0 40
-10
-4.2,'-
Q) k
0 0
0 C) 0
z-f.
-- - G�
- -
- C- � -- -
- - - � C-- -� - -
- - 0 0 -,-. - -0 -
4-- - 0 C - - cC
- - C - -
* C - -
C � o _
o I I I I - I I C* 0 0 0
Cl)
0
z
� z�C/) -
-s Lz.� --� � -r(r� -�
0 � -� ��z C,) --5
Q� �
QOHCl) 0
I
C- �-
C�
oU, -
z0 �
- 7�N� 0
� a* * IIo � H�1-
0 -
7K �4iTH
i-i .� -�--� C.)'I K
�--< 0
- o� 0�v 0.
o 0.I-5-
0)
Q z0 S
- I
�'4
o £
z -
oU,
� Z 00
I-0 -U � �')\S..-.
- / -- II
0 Cu-� "� I> C I
z -
0 C-
0).0 -�
Al
0 -, - AlZ : I - II '-
"or¢ - -"
cj.• .- = -
- II I 6'I
0
z - '- C -,,Y -- l
• - - I.I
- I . -Z
- Z Tit
o - -: - " -. _
zz
- 1.0II
zO - C� £
� -
� Q0 � -U
z
C
0
Co 0 0 C
.7
+I.
- * I -
:- -
2: - £
2: - -� -
2: --
0 � -
Q Cz� 'N - -I- 2 7 *. I-
- -�
0 � - '**�=Cl) c
E-� �.4 CS =
�L) b.� I - -� - .* -� .*
- +'N - � -�
0 '-' - - '- -
�I.- .� -� . �
0 � .- .- �
�0 2- II k � � �
0 - Z U)o� - I I* 0 0
I'
I'
_ / C
�
2:0 4� E-' 0
2:0 � *1
z
o :� .1
C
C
0
- = -
I.zz. -
-r�i) -
z �- -
I.'4 - -
* 'N
00 � = -
-� r- - 'N
Cl) � C 1 'N =-� U) �
'N - -
- - _ -' ZN�Ii
� -� � $- f
� �- � 00 -
� �oI I I Iz
* 0 0
S
0�
- .f.
0
� 0
o � C
0
CC -
- C C -
U2
u-I
(*� 0 -�
�
C � -" INQz � -� - U.2
I -. -C- - - f � e - -
�-I � ..- - -. � - - _
-� C/� IN IN = (Zp2.. � I �
- - -� zZ 0 -= �rI� 7 1-. � -�
- H -� *� �' I.
A � -� � �0 ;.z �- � .� *� � -4-a -
0 -.
0 � � *�
� �2 k E*-� &-. .� -� �z -� o�
I I I0 0 0
S (
7Z C-
1.--C
x ..
1 4.fw T .- w w
- - V)
C- -
zn L -
Q~ 00, U*
z .-
~E F-
top related