- PowerPoint PPT Presentation

!,/&"012,)"'34,&"',%.'5$."6'7/62"88$%&' @ANU ML Workshop, Sept 23, 2011 !"#$%&'($"' )"#$%&*#$"+,%-*".-*,-'

Obama @ Texas

9",)$:;'$%':<$8'6%)$%"'=6/).' Mar’10 : 24 hrs/minute 10% of internet traffic Oct’09: 4 billion photos ~ 12 PB/yr ?? one year of digital life 6000+/minute ~200 GB? ~ 500 TB Apr’09 : 15 billion photos +220 million/week ~ 1.5PB news broadcast ten channels, one year 1,300 GB, 1,830 hrs > ?,%;'6:<"/'"#,4@)"8A' B C%./$6.'D'37<6%"'E?,@':/,2"8' B ?6F$)"'@<6%"'2,))'/"26/.8' B C))'"%G$/6%4"%:,)'H",:-/"8'8"%8".'.,$);'6G"/'C-8:/,)$,I'

J-:)$%"' ' > !,/&"082,)"K'=<;I' B !"88'@/6%"':6'6G"/L:I'M"N"/'&"%"/,)$O,P6%' B Q,:,'G8'$%:"))$&"%2"' > 12,)$%&'-@',%'15?'2),88$L"/' > 7/62"88$%&'26%8$."/,P6%8' > Q,:,0$%:"%8$G"'@/62"88$%&'H/,4"=6/R' B S,.66@'T',@@)$2,P6%8'

Q,:,'IU'76="/' the winning approaches and intervals [Banko and Brill ACL 01] Task: confusion set disambiguation

V#,4@)"A'12"%"'W64@)"P6%' ' “… initial experiments with the GIST descriptor on ten thousand images were very discouraging … however increasing the dataset to one million yielded a qualitative leap.” [Hays and Efros SIGGRAPH07]

?!'1;8:"48'

5$8-,)'W6%2"@:'Q":"2P6%' Task: score each image independently w.r.t. a set of pre-defined visual concepts.

X."46Y' 80% precision, Taxonomy-refined tags @4 tags per image Tagging Precision (%) Normalized classifier tags Precision-calibrated tags “ImageNet-1000”, UIUC-NEC Raw classifier tags (baseline) “Social 20” KNN-voting [Li, Snoek’09] “ImageNet-1000”, KNN “ImageNet-1000”, libLin* Number of tags per image Aggregated performance over 50 “core” visual categories [Xie et al’11].

1-@@6/:'5"2:6/'?,2<$%"8'

15?'[$%'?!'2),88\'' > Z/,$%$%&' Q-,)'H6/4A'C']7' 4,:),F'26."A' alpha_star = quadprog(diag(z)*y'*y*diag (z), -ones(1, Nf), zeros(1, Nf) … ) > Z"8P%&'

^<,:'<;@"/0@,/,4":"/8I' <6='4-2<'4$802),88$L2,P6%I' 76);'%64$%,)'."&/""'.I' =<,:'R"/%")I'

15?'S;@"/0@,/,4":"/'1")"2P6%'

S6=':6'8")"2P6%'W',%.'&,44,I' > 7"/H6/4,%2"'8-/H,2"'=*/*:'<;@"/0@,/,4":"/'<$&<);'%6%0/"&-),/' > c0H6).'2/6880G,)$.,P6%'8""4':6',@@/6#$4,:"':"8:'"//'="))*' XQ-,%'":',)*K'_"-/6264@-P%&'`aabY'

^<"/"K',%.'<6='4-2<':6'8",/2<I' > !$F15?'E-$."' > <N@ADDF)6&*846),*6/&' > _6/4,)$O"';6-/'.,:,' > W"%:"/'&,44,',/6-%.' d-,%P)"'8:,:8'6H'$%G"/8"' .$8:,%2"'$%':<"'.,:,*' > 1",/2<'W'$%')6&'8@,2"*' ]A'=<,:'"G,)-,P6%'2/$:"/$,':6'8",/2<'6%I' ]A'=<,:e8':<"'"//6/'-@@"/'F6-%.'H6/')",G"06%"06-:'2/6880G,)$.,P6%I''

15?A'^<"/"'$8':<"'264@-:,P6%'F6N)"'%"2RI' ]7A'J[_f`\gJ[_fb\' > Z/,$%$%&' > Z"8P%&' J[_f`\'F/-:"0H6/2"' W,%'="'"G"%'8:6/"'c'H6/'),/&"'_I' > 7,/,4":"/'8")"2P6%' "*&*'_Uh"i' C..8'-@'d-$2R);A'j'H6).'#'k'G,)-"8'6H'W'#'b'R"/%")':;@"8'#'j'@,/,4":"/'G,)-"8'l''

E7m'1@"".'-@'6H'c"/%")'W64@-:,P6%' > Z/,$%$%&'8@"".-@'gha#'' B iaaa0.$4"%8$6%'$%@-:'.,:,' B m@':6'baaaT':/,$%$%&'$%8:,%2"8' > Z"8P%&'8@"".-@'II' B 1<6-).'F"'"G"%'46/"' <N@ADD4R),F*$P*&/D@/6n"2:DE7m0!3M15?'

E7m0Z,$)6/".'15?' XW6N"/'":',)K'cQQ'`ahhY' > c";'$.",8' B !6,.T86)G"'=6/R$%&'8":' 6%'E7m'[8$O"'hi\' B W)-8:"/'6%'8@,/8$:;' @,N"/%8':6'$4@/6G"'3DJ' B h 8: '6/."/'<"-/$8P28'H6/' 2<668$%&','=6/R$%&'8":' <N@ADDo2*-2<$2,&6*".-Dg26N"/D@/6n"2:8D&:8G4'

715?'XW<,%&'":*',)*'_371eapY' code.google.com/p/psvm > c";'$.",A' B m8"'3Wq'[$%264@)":"'W<6)"8R;'q,2:6/$O,P6%\':6' ,@@/6#$4,:"'R"/%")'4,:/$#*' B 7"/H6/4'26)-4%0F,8".'3Wq',%.'.$8:/$F-:".'4,:/$#' 4-)P@)$2,P6%':6',2<$"G"'@,/,))")$O,P6%' > gkh#'8@"".'-@'-8$%&'ja'4,2<$%"8' B J%'`aac'$4,&"8'

random subspace bagging Features Training Examples SVM 1 SVM 2 Classifiers [Yan, Tesic and Smith KDD07] 22

Many approaches for scaling up N n n N [Yan et. al. KDD’07] … N p * p Working set on GPU N [Chang et. al. NIPS’07] Large number of models vs. large models ! Some applicable to other models (e.g. graph construction) ! Other issues: normalize input, imbalanced training data, normalize output? ! 23

J:<"/'@,/:8'6H':<"'?!'@$@")$%"'

V#,4@)"'r!,/&"s'7/62"88$%&'Z,8R8' > Z,8R'hA'7CWC!'5JW'W<,))"%&"'`aha' B `a':,/&":'2),88"8' B haKhab':/,$%$%&'$4,&"8K'`bcT'6Fn"2:8' > Z,8R'`A'_31Z'Z9VW'?-)P4".$,'VG"%:'Q":"2P6%' [?VQ\'M"%2<4,/R' B hj'"G"%:8' B hbKaaaT':/,$%$%&'G$."68K'b`KaaaT':"8P%&' 8,4@)".'G$."6'H/,4"8A'iiacT':/,$%$%&K'g`?':"8P%&''

W6)6/'S$8:6&/,4' ! M6o2"))$K'.,'5$%2$K'?6%":K'6/'?$/6I'

5$8-,)'^6/.8'H6/'34,&"'9":/$"G,)' > 562,F-),/;'8$O"'hc'g'h?' > !62,)'.":"2:6/8' B Q6EK'S"88$,%0Ct%"K'S,//$80Ct%"K' ?1V9K'S,//$80W6/%"/K'":2*' > !62,)'."82/$@:6/8' B 13qZK'7WC013qZK'E31ZK'1m9qK'":2*' > Q$u"/"%:'26.$%&'4":<6.8' B c04",%8K'<,8<$%&K'8@,/8"'26.$%&K' ":2'

M,2R06H0:<"0V%G")6@'7),%%$%&' > 1:6/"'/,='.,:,A'giaaEM'G$."68' > q",:-/"'"#:/,2P6%'["*&*'M6^\' B M6^A'`'8"2'D'$4,&"' > 5JW'A'gb'<6-/8v''?VQA'ia'W7m'.,;8'' > 1:6/$%&'/,='13qZ'H",:-/"8A'ghZM'H6/'?VQ':/,$%$%&' B JFn"2:'H",:-/"8A'`a'8"28'D'$4,&"'H6/''`aaT'6Fn"2:8'**'' > wh';",/'W7m'P4"'l''b'46%:<8'6%','d-,.026/"'7W' > !",/%'46.")' B 1:6/"'h?'#'`c''.$48'$%'4,:),FA'hi*kEMx' B 15?'gh'4$%'@"/']7'y'zW5[j\'y'z'@,/,4'[`k\' ' " '`'<6-/8x'

> Z<"'&66.'%"=8' B ?68:'6H':<"8"'264@-:,P6%'$8'.,:,0',%.':,8R0' @,//,)$O,F)"' B 1:6/"'$%:"/4".$,:"'/"8-):8' B 4,%;'84,))'L)"8'<,/4H-)'[x\' > r:,/'2HOsv'<,8<$%&'<")@8'' r4,2<$%"')",/%$%&s' " ' > ^<,:'$H'l' B 7W'.$"8I' B 76="/'6-:,&"I' B _":=6/R'P4"6-:I''

HBase Pig Hive Chukwa Zoo MapReduce HDFS Keeper Core Avro

?,@9".-2"'

q,2"F66R'"#,4@)"A'2<"2R0$%'P4"8' <N@ADD===*H,2"F66R*264D%6:"8DH,2"F66R0"%&$%""/$%&D$%:"/%$%&0,:0H,2"F66R0=<60 &6"80=<"/"0=<"%0,%.0=<;0$:04,N"/8Dhahja`{`i{h|i{|`a'

1-44,/;' > !,/&"'.,:,'$8':<"':/"%.' > ?!',)&6/$:<48' " '@/,2P2"'' > C@@)$2,P6%08@"2$L2'@),%%$%&'$8'$4@6/:,%:' > ?,%;'@,/,))")'264@-P%&'@,/,.$&48':6'<")@'' > 6:<"/':6@$28'l' B !,/&"',46-%:8'6H'8@,/8"'H",:-/"8' B E??K'W9q',%.'6:<"/'46.")8' B 3%G"/:".'$%."#K'<,8<$%&K'$%."#$%&'**''

Z<,%R8x']-"8P6%8I' > 1)$."'2/".$:8' B S,8P"'F66R'rZ<"'V)"4"%:8'6H'1:,P8P2,)'?,2<$%"'!",/%$%&s'' B ^$%8:6%'S8-'[_Zm'Z,$=,%\K'96%&'},%'[q,2"F66R\' B W3c?':-:6/$,)'F;'V.=,/.'W<,%&'[E66&)"\' B ^^^ehh':-:6/$,)'F;'C)"#'146),'r&/,@<$2,)'46.")8'H6/':<"'$%:"/%":s'' B q)$2R/'-8"/'+;-)$;,/:' > J:<"/'@6$%:"/8' B ~6<%'!,%&H6/.'),/&"082,)"':-:6/$,)+'cQQehh''