Parallel Sequential Minimal Optimization for the Training. of Support Vector Machines

Size: px
Start display at page:

Download "Parallel Sequential Minimal Optimization for the Training. of Support Vector Machines"

Transcription

1 Parallel Sequental Mnmal Optmzaton for the Tranng of Sport Vector Machnes 1 L.J. Cao a, S.S. Keerth b, C.J. Ong b, P. Uvaraj c, X.J. Fu c and H.P. Lee c, J.Q. Zhang a a Fnancal Studes of Fudan Unversty, HanDan Road, ShangHa, P.R. Chna, b Dept. Of Mechancal Engneerng, Natonal Unversty of Sngapore, 10 Kent Rdge Crescent, Sngapore c Insttute of Hgh Performance Computng, 1 Scence Par Road, #01-01 the Caprcorn, Scence Par II, Sngapore Abstract Sequental mnmal optmzaton (SMO) s one popular algorthm for tranng sport vector machne (SVM), but t stll requres a large amount of computaton tme for solvng large sze problems. Ths paper proposes one parallel mplementaton of SMO for tranng SVM. The parallel SMO s developed usng message passng nterface (MPI). Specfcally, the parallel SMO frst parttons the entre tranng data set nto smaller subsets and then smultaneously runs multple CPU processors to deal wth each of the parttoned data sets. Experments show that there s great speed on the adult data set and the MNIST data set when many processors are used. There are also satsfactory results on the Web data set. Index Terms Sport vector machne (SVM), sequental mnmal optmzaton (SMO), message passng nterface (MPI), parallel algorthm 1 Correspondng author. Emal: ljcao@fudan.edu.cn. The research wor s funded by Natonal Natural Scence Research Fund No and sponsored by Shangha Pujang program. 1

2 I. INTRODUCTION Recently, a lot of research wor has been done on sport vector machnes (SVMs), manly due to ther mpressve generalzaton performance n solvng varous l machne learnng problems [1,,3,4,5]. Gven a set of data ponts { } ( ( X, ) d X R s the nput vector of th tranng data pattern; y { 1, 1} s ts class label; y l s the total number of tranng data patterns), tranng an SVM n classfcaton s equvalent to solvng the folng lnearly constraned convex quadratc programmng (QP) problem. maxmze: l l l 1 R( α ) = α αα j y y j( X, X j ) (1) = 1 = 1 j= 1 l subject to: α y = 0 () = 1 0 α c, = 1, L, l where ( X, X j ) s the ernel functon. The mostly wdely used ernel functon s the Gaussan functon e X X j σ, where σ s the wdth of the Gaussan ernel. α s the Lagrange multpler to be optmzed. For each of tranng data patterns, one α s assocated. c s the regularzaton constant pre-determned by users. After solvng the QP problem (1), the folng decson functon s used to determne the class label for a new data pattern. l functon( X ) = α y( X, X ) + b (3) =1 where b s obtaned from the soluton of (1). So the man problem n SVM s reduced to solvng the QP problem (1), where the number of varables α to be optmzed s equal to the number of tranng data

3 patterns l. For small sze problems, standard QP technques such as the projected conjugate gradent can be drectly appled. But for large sze problems, standard QP technques are not useful as they requre a large amount of computer memory to store the ernel matrx K as the number of elements of K s equal to the square of the number of tranng data patterns. For mang SVM more practcal, specal algorthms are developed, such as Vapn s chunng [6], Osuna s decomposton [7] and Joachms s SVM lght [8]. They mae the tranng of SVM possble by breang the large QP problem (1) nto a seres of smaller QP problems and optmzng only a subset of tranng data patterns at each step. The subset of tranng data patterns optmzed at each step s called the worng set. Thus, these approaches are categorzed as the worng set methods. Based on the dea of the worng set methods, Platt [9] proposed the sequental mnmal optmzaton (SMO) algorthm whch selects the sze of the worng set as two and uses a smple analytcal approach to solve the reduced smaller QP problems. There are some heurstcs used for choosng two α to optmze at each step. As ponted out by Platt, SMO scales only quadratcally n the number of tranng data patterns, whle other algorthms scales cubcally or more n the number of tranng data patterns. Later, Keerth et. al. [10,11] ascertaned neffcency assocated wth Platt s SMO and suggested two modfed versons of SMO that are much more effcent than Platt s orgnal SMO. The second modfcaton s partcular good and used n popular SVM pacages such as LIBSVM [1]. We wll refer to ths modfcaton as the modfed SMO algorthm. Recently, there are few wors on developng parallel mplementaton of tranng SVMs [13,14,15,16]. In [13], a mxture of SVMs are traned n parallel usng the subsets of a tranng data set. The results of each SVM are then combned by tranng 3

4 another mult-layer perceptron. The experment shows that the proposed parallel algorthm can provde much effcency than usng a sngle SVM. In the algorthm proposed by Dong et. al. [14], multple SVMs are also developed usng subsets of a tranng data set. The sport vectors n each SVM are then collected to tran another new SVM. The experment demonstrates much effcency of the algorthm. Zanghrat and Zann [15] also proposed a parallel mplementaton of SVM lght where the whole quadratc programmng problem s splt nto smaller subproblems. The subproblems are then solved by a varable projecton method. The results show that the approach s comparable on scalar machnes wth a wdely used technque and can acheve good effcency and scalablty on a multprocessor system. Huang et. Al. [16] proposed a modular networ mplementaton for SVM. The result found out that the modular networ could sgnfcantly reduce the learnng tme of SVM algorthms wthout sacrfcng much generalzaton performance. Ths paper proposes a parallel mplementaton of the modfed SMO based on the multprocessor system for speedng the tranng of SVM, especally wth the am of solvng large sze problems. In ths paper, the parallel SMO s developed usng message passng nterface (MPI) [17]. Unle the sequental SMO whch handles the entre tranng data set usng a sngle CPU processor, the parallel SMO frst parttons the entre tranng data set nto smaller subsets and then smultaneously runs multple CPU processors to deal wth each of the parttoned data sets. On the adult data set the parallell SMO usng 3 CPU processors s more than 1 tmes faster than the sequental SMO. On the web data set,the parallel SMO usng 30 CPU processors s more than 10 tmes faster than the sequental SMO. On the MNIST data set the parallel SMO usng 30 CPU processors on the averaged tme of one-aganst-all SVM classfers s more than 1 tmes faster than the sequental SMO. 4

5 Ths paper s organzed as fols. Secton II gves an overvew of the modfed SMO. Secton III descrbes the parallel SMO developed usng MPI. Secton IV presents the experment ndcatng the effcency of the parallel SMO. A short concluson then fols. I I II. A BRIEF OVERVIEW OF THE MODIFIED SMO We begn the descrpton of the modfed SMO by gvng the notaton used. Let = { : y = 1, 0 < α < c} { : y = 1, 0 < α }, I { : = 1, α 0}, 0 < c 1 = y = = { : y = 1, α }, I = { : y = 1, α }, and I { : = 1, α 0}. = c 3 = c 4 = y = I = U I, = 0, L, 4 denotes the ndex of tranng data patterns. f l = α j y j( X j, X ) y. b = mn{ f : I 0 I1 I }, I = arg mn f. j=1 b =. τ =10 6. = max{ f : I 0 I 3 I 4}, I arg max f The dea of the modfed SMO s to optmze the two α assocated wth b and b accordng to (4) and (5) at each step. Ther assocated ndex are I and I. α α new new 1 old old y ( f f ) old 1 = α (4) η old old new = α + s α α ) (5) 1 ( where the varables assocated wth the two α are represented usng the subscrpts 1 and. s = y y. (, ), ) (, ) new 1 η = X 1 X (X 1 X 1 X X. α 1 and new need to be clpped to [ 0, C]. That s, 0 α c and 0 α new c. 1 new α After optmzng α 1 and α, f, denotng the error on the th tranng data pattern, s dated accordng to the folng: f new old new old new old = f + α α ) y ( X, X ) + ( α α ) y ( X, X ) (6) (

6 Based on the dated values of f, b and b and the assocated ndex I and I are dated agan accordng to ther defntons. The dated values are then used to choose another two new α to optmze at the next step. In addton, the value of Eq. (1), represented by Dual, s dated at each step Dual new α α new old new old old 1 1 old old 1 1 = Dual ( f1 f ) + η( ) (7) y1 y1 1 α α And DualtyGap, representng the dfference between the prmal and the dual objectve functon n SVM, s calculated by (8). DualtyGap = l = 0 α y f + l = o ε (8) ε = Cmax(0, where ε = Cmax(0, b - f - b ) + f ) f f y y = 1 = 1 A more detaled descrpton of Dual and DualtyGap can be referred to the paper [8]. Dual and DualtyGap are used for checng the convergence of the program. A smple descrpton of the modfed SMO n the sequental form can be summarzed as: Sequental SMO Algorthm: Intalze α = 0, f =, Dual = 0, = 1, L, l y Calculate b, I, b, I, DualtyGap Untl DualtyGap τ Dual (1) Optmze α I, α I () Update f, = 1, L, l (3) Calculate b, I, b, I, DualtyGap and date Dual Repeat 6

7 III. THE PARALLEL SMO MPI s not a new programmng language, but a lbrary of functons that can be used n C, C++ and FORTRAN [17]. MPI als one to easly mplement an algorthm n parallel by runnng multple CPU processors for mprovng effcency. The Sngle Program Multple Data (SPMD) mode where dfferent processors execute the same program but dfferent data s generally used n MPI for developng parallel programs. In the sequental SMO algorthm, most of computaton tme s domnated by datng f array at the teraton (), as t ncludes the ernel evaluatons and s also requred for every tranng data pattern. As shown n our experment, over 90% of the total computaton tme of the sequental SMO s used for datng f array. So the frst dea for us to mprove the effcency of SMO s to develop the parallel program f for datng array. Accordng to (6), datng array s ndependently evaluated one tranng data pattern at a tme, so the SPMD mode can be used to execute ths program n parallel. That s, the entre tranng data set s frstly equally parttoned nto smaller subsets accordng to the number of processors used. Then each of the parttoned subsets s dstrbuted nto one CPU processor. By executng the program f of datng f array usng all the processors, each processor wll date a dfferent subset of f array based on ts assgned tranng data patterns. In such a way, much computaton tme could be saved. Let p denotes the total number of processors used, t f s the amount of computaton tme used for datng f array n the sequental SMO. By usng the parallel program of datng f array, the amount of computaton tme used to date f array s almost reduced to 1 t f p. 7

8 f Besdes datng array, calculatng,, and can also be performed n parallel as the calculaton nvolves examnng all the tranng data ponts. By executng the program of calculatng,, and usng all the processors, each processor could obtan one b and one b as well as the assocated b b b b I I I I I and I based on ts assgned tranng data patterns. The b,, and I b I of each processor are not global n the sense they are obtaned only based on a subset of all the tranng data patterns. The global and global b are respectvely the mnmum value of b of each processor and the maxmum value of of each b b processor, as descrbed n Secton. By determnng the global b and the global b, the assocated I and I can thus be found out. The correspondng two α are then optmzed by usng any one CPU processor. Accordng to (8), calculatng DualtyGap s also ndependently evaluated one tranng data pattern at a tme. So ths program can also be executed n parallel usng the SPMD mode. By runnng the program of Eq. (8) usng multple CPU processors, each processor wll calculate a dfferent subset of DualtyGap based on ts assgned tranng data patterns. The value of DualtyGap on the entre tranng data patterns s the sum of the DualtyGap of all the processors. In summary, based on the SPMD parallel mode, the parallel SMO date F array and calculate b, b, I, I, and DualtyGap at each step n parallel usng multple CPU processor. The calculaton of other parts of SMO whch tae lttle tme s done usng one CPU processor, whch s the same as used n the sequental SMO. Due to the use of multple processors, communcaton among processors s also requred n the parallel SMO, such as gettng global b, I, b and I from, b 8

9 I, and of each processor. For mang the parallel SMO effcent, the b I communcaton tme should be ept small. A bref descrpton of executng the parallel SMO can be summarzed as fols. Parallel SMO Algorthm: p Notaton: p s the total number of processors used. }, U l = l s a subset of { l = 1 =1 p all the tranng data patterns and assgned to processor., b, I,,, f b I DualtyGap,, l denote the varables assocated wth processor. α f l = α j=1 j y j (X j, X ) y. I1 I = mn{ f : I 0 l }, I = arg mn f. b b = max{ f : I 0 I 3 I l }, I = arg max f. b, I, b,, and 4 I DualtyGap stll denote the varables on the entre tranng data patterns. b = max{ b } b, I = arg b =, b = max{ }, = arg b = b, I b I I DualtyGap = p = 1 DualtyGap. Intalze α = 0, f = y, Dual = 0, l, Calculate,,,, b I b I DualtyGap = 1, L, p Obtan b, I, b, I, and DualtyGap Untl DualtyGap τ Dual (1) Optmze α α I, I () Update f, l (3) Calculate b,,,, I b I DualtyGap 9

10 Repeat (4) Obtan b, I, b, I, DualtyGap and date Dual A more detaled descrpton of the parallel SMO can be referred to the pseudocode n appendx A. IV. EXPERIMENT The parallel SMO s tested aganst the sequental SMO usng three benchmars: the adult data set, the web data set and the MNIST data set. Both algorthms are wrtten n C. Both algorthms are run on IBM p690 Regata SerComputer whch has a total of 7 nodes, wth each node havng 3 power PC_POWER4 1.3GHz processors. For ensurng the same accuracy n the sequental SMO and the parallel SMO, the stop crtera used n both algorthms such as the value of τ are all the same. A. Adult Data Set The frst data set used to test the parallel SMO s speed s the UCI adult data set [10]. The tas s to predct whether the household has an ncome larger than $50,000 based on a total of 13 bnary attrbutes. For each nput vector, only an average of 14 bnary attrbutes are true, represented by the value of 1. Other attrbutes are all false, represented by the value of 0. There are a total of 8,956 data patterns n the tranng data set. The Gaussan ernel s used for both the sequental SMO and the parallel SMO. The values of Gaussan varance σ and c are arbtrarly used as 100 and 1. These values are not necessarly ones that gve the best generalzaton performance of SVM, as the purpose of ths experment s only for evaluatng the computaton tme of two 10

11 algorthms. Moreover, the LIBSVM verson.8 proposed by Chang and Ln [1] s also nvestgated usng a sngle processor on the experment. The am s to see whether the ernel cache used n LIBSVM can provde effcency n comparson wth the sequental SMO wthout ernel cache. The elapsed tme (measured n seconds) wth dfferent number of processors n the sequental SMO, the parallel SMO and LIBSVM s gven n Table 1, as well as the number of converged sport vectors (denoted as SVs) and bounded sport vectors wth α = c ( denoted as BSVs). From the table, t can be observed that the elapsed tme of the parallel SMO gradually reduces wth an ncrease n the number of processors. It can be reduced by almost half wth the use of two processors and almost three-quarters wth the use of four processors, etc.. Ths result demonstrates that the parallel SMO s effcent n reducng the tranng tme of SVM. Moreover, the parallel SMO usng one CPU processor taes slghtly more tme than the sequental SMO, due to the use of MPI programs. The table also shows that LIBSVM runnng on the sngle processor requres less tme than that of the sequental SMO. Ths demonstrates that the ernel cachng s effectve n reducng the computaton tme of the ernel evaluaton. For evaluatng the performance of the parallel SMO, the folng two crtera are used: speed and effcency. They are respectvely defned by the elapsed tme of the sequental SMO speed = (9) the elapsed tme of the parallelsmo effceny = speed number of processors (10) The speed of the parallel SMO wth respect to dfferent number of processors s llustrated n Fg. 1. The fgure shows that to 16 processors the parallel SMO scales almost lnearly wth the number of processors. After that, the scalablty of the parallel 11

12 SMO s slghtly reduced. The maxmum value of the speed s more than 1, correspondng to the use of 3 processors. The result means that the tranng tme of 1 the parallel SMO by runnng 3 processors s only about of that of the sequental 1 SMO, whch s very good. The effcency of the parallel SMO wth dfferent number of processors s llustrated n Fg.. As shown n the fgure, the value of the effcency of the parallel SMO s when two processors are used. It gradually reduces as the number of processor ncreases. The reason may le n that the use of more processors wll lead to more communcaton tme, thus reducng the effcency of the parallel SMO. For a better understandng of the cost of varous subparts n the parallel SMO, the computaton tme n dfferent components (I/O; ntalzaton; optmzng α I and α ; datng and calculatng,,,, DualtyGap ; and obtanng I f b I b I b, I, b, I, DualtyGap ) s reported n Table. The tme for datng f and calculatng,,,, DualtyGap s called as the parallel tme as the b nvolved calculatons are done n parallel. And the tme for obtanng b, I, b, I I b I, DualtyGap s called as the communcaton tme as there are many processors ncluded n the calculaton. The table shows that the tme for I/O, ntalzaton, and optmzng α I and α I s lttle and rrelevant to the number of processor, whle a large amount of tme s used n the parallel tme, whch means that the datng of f and the calculatng of b, I, b, l I ow, DualtyGap had better be performed n parallel usng multple processors. As expected, the parallel tme decreases wth the ncrease of the number of processors. In contrast, the communcaton tme slghtly 1

13 ncreases wth the ncrease of the number of processors. Ths exactly explans why the effcency of the parallel SMO decreases as the number of processors ncreases. B. Web Data Set The web data set s examned n the second experment [10]. Ths problem s to classfy whether a web page belongs to a certan category or not. There are a total of 4,69 data patterns n the tranng data set, wth each data pattern composed of 300 spare bnary eyword attrbutes extracted from each web page. For ths data set, the Gaussan functon s stll used as the ernel functon of the sequental SMO and the parallel SMO. The values of Gaussan varance are respectvely used as and 64. σ and c The elapsed tme wth dfferent number of processors used n the sequental SMO, the parallel SMO and LIBSVM s gven n Table 3, as well as the total number of sport vectors and bounded sport vectors. Same as n the adult data set, the elapsed tme of the parallel SMO gradually reduces wth the ncrease of the number of processors, by almost half usng two processors and almost three-quarters usng four processors, so on and so for. The parallel SMO usng one CPU processor also taes slghtly more tme than the sequental SMO, due to the use of MPI program. The LIBSVM requres less tme than that of the sequental SMO, due to the use of the ernel cache. Based on the obtaned results, the speed and the effcency of the parallel SMO are calculated and respectvely llustrated n Fg. 3 and Fg. 4. Fg. 3 shows that the speed of the parallel SMO ncreases wth the ncrease of the number of processors ( to 30 processors), demonstratng the effcency of the parallel SMO. For ths data set, the maxmum value of the speed s more than 10, correspondng to the use of 13

14 30 processors. As llustrated n Fg. 4, the effcency of the parallel SMO decreases wth the ncrease of the number of processors, due to the ncrease of the communcaton tme. The computaton tme n dfferent components of the parallel SMO s reported n Table 4. The same conclusons are reached as n the adult data set. The tme for I/O, ntalzaton, and optmzng α I and I α s lttle and almost rrelevant to the number of processors. Wth the ncrease of the number of processors, the parallel tme decreases, whle the communcaton tme slghtly ncreases. In terms of speed and effcency the result on the web data set s not as good as that n the adult data set. Ths can be analyzed as the rato of the parallel tme to the communcaton tme n the web data set s much smaller than that of the adult data set, as llustrated n Table and Table 4. Ths also means that the advantage of usng the parallel SMO s more obvous for large sze problems. C. MNIST Data Set The MNIST handwrtten dgt data set s also examned n the experment. Ths data set conssts of 60,000 tranng samples and 10,000 testng samples. Each sample s composed of 576 features. Ths data set s avalable at and has also been used n Dong et al. s wor on speedng the sequental SMO [18]. The MNIST data set s actually a ten-class classfcaton problem. Accordng to the one aganst the rest method, ten SVM classfers are constructed by separatng one class from the rest. In our experment, the Gaussan ernel s used n the sequental SMO and the parallel SMO for each of ten SVM classfers. The values of σ and c are respectvely used as 0.6 and 10, same as those used n [14]. 14

15 The elapsed tme wth dfferent number of processors n the sequental SMO and the parallel SMO and LIBSVM for each of ten SVM classfers s gven n Table 5. The number of converged sport vectors and bounded sport vectors s descrbed n Table 6. The averaged value of the elapsed tme n the ten SVM classfers s also lsted n ths table. The table shows that there s stll beneft n the usng of the ernel cache n LIBSVM n comparson wth the sequental SMO. Fg. 5 and Fg. 6 respectvely llustrate the speed and the effcency of the parallel SMO. Fg. 5 shows that the speed of the parallel SMO ncreases wth the ncrease of the number of processors. The maxmum values of the speed n the ten SVM classfers range from 17.1 to.8. The averaged maxmum value of speed s equal to 1.7, correspondng to the use of 30 processors. Fg. 6 shows that the effcency of the parallel SMO decreases wth the ncrease of the number of processors, due to the use of more communcaton tme. V. CONCLUSIONS Ths paper proposes the parallel mplementaton of SMO usng MPI. The parallel SMO uses multple CPU processors to deal wth the computaton of SMO. By parttonng the entre tranng data set nto smaller subsets and dstrbutng each of the parttoned subsets nto one CPU processor, the parallel SMO dates F array and calculates b, b, and DualtyGap at each step n parallel usng multple CPU processors. Ths parallel mode s called the SPMD model n MPI. Experment on three large data sets demonstrates the effcency of the parallel SMO. The experment also shows that the effcency of the parallel SMO decreases wth the ncrease of the number of processors, as there s more communcaton tme wth 15

16 the use of more processors. For ths reason, the parallel SMO s more useful for large sze problems. The experment also shows that LIBSVM wth the usng of the worng set sze as s more effcent than the sequental SMO. Ths can be explaned that the LIBSVM use the ernel cache, whle the sequental and parallel SMO do not tae t nto account. Future wor wll explot the ernel cache for further mprovng the current verson of the parallel SMO. In the current verson of the parallel SMO, the mult-class classfcaton problem s performed by consderng one class by one class. In the future wor, t s worthy to perform the mult-class classfcaton problem n parallel by consderng all the classes smultaneously for further mprovng the effcency of the parallel SMO. In such an approach, t needs to develop a structural approach to consder the communcaton between processors Ths wor s very useful for the research where multple CPU processors machne s avalable. Future wor also needs to extend the parallel SMO from classfcaton for regresson estmaton by mplementng the same methodology for SVM regressor. 16

17 References: [1] V.N. Vapn, The Nature of Statstcal Learnng Theory, New Yor, Sprnger- Verlag, [] C.J.C. Burges, A tutoral on sport vector machnes for pattern recognton, Knowledge Dscovery and Data Mnng, Vol., No., pp , [3] L.J. Cao and F. E.H. Tay, Sport vector machnes wth adaptve parameters n fnancal tme seres forecastng, IEEE Transactons on Neural Networs, 14(6), ,003. [4] S. Gutta, R.J. Jeffrey, P. Jonathon and H. Wechsler, Mxture of Experts for Classfcaton of Gender, Ethnc Orgn, and Pose of Human Faces, IEEE Transactons on Neural Networs, 11 (4), July 000, [5] K. Ieda, Effects of Kernel Functon on Nu Sport Vector Machnes n Extreme Cases, IEEE Transactons on Neural Networs, 17 (1), Jan.006, 1-9. [6] V.N. Vapn, Estmaton of Dependence Based on Emprcal Data, New Yor: Sprnger Verlag, 198. [7] E. Osuna, R. Freund and F. Gros, An mproved algorthm for sport vector machnes, NNSP 97: Proc. of the IEEE Sgnal Processng Socety Worshop, Amela Island, USA, pp , [8] T. Joachms, Mang large-scale sport vector machne learnng practcal, n Advances n Kernel Methods: Sport Vector Machnes, ed. by B. Scholopf, C. Burges, A. Smola. MIT Press, Cambrdge, MA, December [9] J.C. Platt, Fast tranng of sport vector machnes usng sequental mnmal optmsaton, In Advances n Kernel Methods Sport Vector Learnng, ed. by B. Scholopf, C.J.C. Burges and A.J. Smola, pp , MIT Press,

18 [10] S.S. Keerth, S.K. Shevade, C. Bhattaacharyya and K.R.K. Murthy, Improvements to Platt s SMO algorthm for SVM classfer desgn, Neural Computaton, Vol. 13, pp , 001. [11] S.K. Shevade, S.S. Keerth, C. Bhattacharyya and K.R.K. Murthy, Improvements to the SMO algorthm for SVM regresson, IEEE Transactons on Neural Networs, 11 (5), Sept. 000 Page(s): [1] C.C. Chang and C.J. Ln. LIBSVM: a Lbrary for Sport Vector Machnes, avalable at [13] R. Collobert, S. Bengo and Y. Bengo, A parallel mxture of SVMs for very large scale problems, Neural Computaton, Vol. 14, No. 5, pp , 00. [14] J. X. Dong, A. Krzyza, C. Y. Suen, A fast Parallel Optmzaton for Tranng Sport Vector Machne, Proc. of 3rd Int. Conf. Machne Learnng and Data Mnng, P. Perner and A. Rosenfeld (Eds.) Sprnger Lecture Notes n Artfcal Intellgence (LNAI 734), pp , Lepzg, Germany, July 5-7, 003 [15] G. Zanghrat, L. Zann, A parallel solver for large quadratc programs n tranng sport vector machnes, Parallel Computng, Vol. 9, No. 4, pp , 003. [16] B.H. Guang, K. Z. Mao, C.K. Sew and D.S. Huang, Fast Modular Networ Implementaton for Sport Vector Machnes, IEEE Transactons on Neural Networs, Vol. 16, No. 6, Nov. 005, [17] P.S. Pacheco, Parallel Programmng wth MPI, San Francsco, Calf.: Morgan Kaufmann Publshers, [18] J.X. Dong, A. Krzyza and C.Y. Suen, A fast SVM tranng algorthm, accepted n Pattern Recognton and Artfcal Intellgence,

19 Appendx A: Pseudo-code for the parallel SMO ( Note: If there s some process ran before the code, ths means that only the processor assocated wth the ran executes the code. Otherwse, all the processors execute the code. ) n_sample = total number of tranng samples p = total number of processors local_nsample = n_sample/ p Procedure taestep ( ) f ( _==_&& Z1==Z ) return 0; s=y1*y; f ( y1==y ) gamma=alph1+alph; else gamma=alph1-alph; f ( s==1 ) { f (y==1) { L=MAX( 0,gamma-C); H=MIN(C, gamma); } else { L=MAX(0,gamma-C); H=MIN(C, gamma); } } else { L=MAX(0,-gamma); f (y==1) H=MIN(C, C-gamma); else 19

20 H=MIN(C, C-gamma); } f (H<=L) return 0; K11 = ernel ( X1, X1 ); K = ernel ( X, X ); K1 = ernel ( X1, X ); eta=*k1-k11-k; f ( eta<eps*(k11+k) ) { a= alph-(y*(f1-f)/eta); f (a<l) a=l; else f (a>h) a=h; } else { slope=y*(f1-f); change=slope *(H-L); f( fabs(change)>0 ) { f (slope>0 ) a=h; else a=l; } else a=alph; } f (y==1) { f (a> C-EPS*C) a=c; else f (a<eps*c) a=0; else ; } else { f (a>c-eps*c) 0

21 a=c; else f (a<eps*c) a=0; else ; } f( fabs(a-alph)<eps* (a+alph+eps) return 0; f ( s==1 ) a1=gamma-a else a1=gamma+a; f (y1==1) { f (a1> C-EPS*C) a1=c; else f (a1<eps*c) a1=0; else ; } else { f (a1>c-eps*c) a1=c; else f (a1<eps*c) a1=0; else ; } date the value of Dual return 1 Endprocedure Procedure ComputeDualtyGap( ) DualtyGap=0; loop over local_nsample tranng samples f ( y[]==1 ) DualtyGap += C*MAX(0, (b-fcache[]) ); else DualtyGap +=C*MAX(0, (-b+fcache[]) ); 1

22 loop over tranng samples n I_0 and I_ and I_3 DualtyGap+=alpha[]*y[]*fcache[]; return DualtyGap; Endprocedure Procedure Man( ) processor 0: read the frst bloc of local_nsample tranng data patterns from the data fle and save them nto the matrx X for =1 to p read the th bloc of local_nsample tranng data patterns from the data fle and send them to processor end processors 1 to p: receve local_nsample tranng data patterns from processor 0 and save them nto the matrx X (all the processors) ntalze alpha array to all zero (for local_nsample tranng data patterns ) ntalze fcache array to the negatve of y array (for local_nsample tranng data patterns ) store the ndces of postve class n I_1 and negatve class n I_4 (for local_nsample tranng data patterns ) set b to zero ntalze the value of Dual to zero DualtyGap=ComputeDualtyGap( ) (for local_nsample tranng data patterns ) sum DualtyGap of each processor and broadcast t to every processor compute ( b_, _ ) and ( b_, _) usng n I and fcache array (for local_nsample tranng data patterns ) compute global b_ and global b_ usng local b_ and local b_ of each processor fnd out processor Z1 contanng global b_ fnd out processor Z contanng global b_ processor Z1: alph1=alpha[ _ ]; y1=y[ _ ];

23 F1=fcache[ _ ]; X1=X[ _ ]; broadcast alph1, y1, F1, and X1 to every processor processor Z: alph=alpha[ _ ]; y=y[ _]; F=fcache[ _]; X=X[ _]; broadcast alph, y, F, and X to every processor numchanged=1; whle ( DualtyGap>tol*abs(Dual) && numchanged!=0 ) { processor 0: numchanged=taestep( ); broadcast numchanged to every processor f ( numchanged==1 ) { processor 0: broadcast a1, a, and Dual to every processor processor Z1: alph[_ ]=a1; f (y1==1) { f ( a1==c ) move 1 to I_3; else f (a1 ==0 ) move 1 to I_1; else move 1 to I_0; } else { f ( a1==c ) move 1 to I_; else f ( a1==0 ) 3

24 } else move 1 to I_4; move 1 to I_0; processor Z: alph[_]=a; f (y==1) { f ( a==c ) move to I_3; else f ( a==0 ) move to I_1; else move to I_0; } else { f ( a==c ) move to I_; else f (a==0 ) move to I_4; else move to I_0; } (all the processors) date fcache[] for n I usng new Lagrange multplers (for local_nsample tranng data patterns ) compute (b_, _) and (b_, _) usng n I and fcache array (for local_nsample tranng data patterns ) compute global b_ and global b_ usng local b_ and local b_ of each processor fnd out processor Z1 contanng global b_ fnd out processor Z contanng global b_ 4

25 processor Z1: alph1=alpha[ _ ]; y1=y[ _ ]; F1=fcache[ _ ]; X1=X[ _ ]; broadcast alph1, y1, F1, and X1 to every processor processor Z: alph=alpha[ _ ]; y=y[ _]; F=fcache[ _]; X=X[ _]; broadcast alph, y, F, and X to every processor b=(b+b)/ DualtyGap=ComputeDualtyGap( ) sum DualtyGap of each processor and broadcast t to every processor } ( end of whle loop) Endprocedure b=(b+b)/ DualtyGap=ComputeDualtyGap( ) sum DualtyGap of each processor and broadcast t to every processor Prmal=Dual+DualtyGap 5

26 Fg. 1. The speed of the parallel SMO on the adult data set. Fg. 3. The speed of the paralleled SMO on the web data set. Fg.. The effcency of the parallel SMO on the adult data set. 6

27 Fg. 3. The speed of the parallel SMO on the web data set. Fg. 4. The effcency of the parallel SMO on the web data set. 7

28 Fg. 5. The speed of the parallel SMO on the MNIST data set. Fg. 6. The effcency of the parallel SMO on the MNIST data set. 8

29 TABLE I THE ELAPSED TIME (SECONDS) USED IN THE SEQUENTIAL SMO AND THE PARALLEL SMO AND LIBSVM ON THE ADULT DATA SET. LIBSVM Sequental Parallel SMO SMO 1P P 4P 8P 16P 3P Tme(s) SVs BSVs TABLE II THE COMPUATION TIME IN DIFFERENT COMPONENTS OF THE PARALLEL SMO ON THE ADULT DATA SET. Components Number of processors 1P P 4P 8P 16P 3P I/O ntalzaton a I_, a I_ b, I, b, I, DualtyGap F, b, I, b, I, DualtyGap

30 TABLE III THE ELAPSED TIME USED IN THE SEQUENTIAL SMO AND THE PARALLEL SMO AND LIBSVM ON THE WEB DATA SET. LIBSV Sequental Parallel SMO M SMO 1P P 4P 8P 16P 30P Tme(s) SVs BSVs TABLE IV THE COMPUATION TIME IN DIFFERENT COMPONENTS OF THE PARALLEL SMO ON THE WEB DATA SET. Components Number of processors 1P P 4P 8P 16P 30P I/O ntalzaton a I_, a I_ b, I, b, I, DualtyGap F, b, I, b, I, DualtyGap

31 TABLE V THE ELAPSED TIME USED IN THE SEQUENTIAL SMO AND THE PARALLEL SMO AND LIBSVM ON THE MNIST DATA SET. Class LIBSVM Sequental Parallel SMO SMO 1P P 4P 8P 16P 30P Averaged

32 TABLE V THE NUMBER OF CONVERGED SUPPORT VECTORS AND BOUNDED SUPPORT VECTORS IN THE SEQUENTIAL SMO AND THE PARALLEL SMO AND LIBSVM ON THE MNIST DATA SET. Class LIBSVM #SVs #BSVs Sequental SMO Parallel SMO 1P P 4P 8P 16P 30P #SVs #BSVs #SVs #BSVs #SVs #BSVs #SVs #BSVs #SVs #BSVs #SVs #BSVs #SVs #BSVs Averaged

Support Vector Machines

Support Vector Machines /9/207 MIST.6060 Busness Intellgence and Data Mnng What are Support Vector Machnes? Support Vector Machnes Support Vector Machnes (SVMs) are supervsed learnng technques that analyze data and recognze patterns.

More information

Learning the Kernel Parameters in Kernel Minimum Distance Classifier

Learning the Kernel Parameters in Kernel Minimum Distance Classifier Learnng the Kernel Parameters n Kernel Mnmum Dstance Classfer Daoqang Zhang 1,, Songcan Chen and Zh-Hua Zhou 1* 1 Natonal Laboratory for Novel Software Technology Nanjng Unversty, Nanjng 193, Chna Department

More information

Parallelism for Nested Loops with Non-uniform and Flow Dependences

Parallelism for Nested Loops with Non-uniform and Flow Dependences Parallelsm for Nested Loops wth Non-unform and Flow Dependences Sam-Jn Jeong Dept. of Informaton & Communcaton Engneerng, Cheonan Unversty, 5, Anseo-dong, Cheonan, Chungnam, 330-80, Korea. seong@cheonan.ac.kr

More information

Support Vector Machines. CS534 - Machine Learning

Support Vector Machines. CS534 - Machine Learning Support Vector Machnes CS534 - Machne Learnng Perceptron Revsted: Lnear Separators Bnar classfcaton can be veed as the task of separatng classes n feature space: b > 0 b 0 b < 0 f() sgn( b) Lnear Separators

More information

Course Introduction. Algorithm 8/31/2017. COSC 320 Advanced Data Structures and Algorithms. COSC 320 Advanced Data Structures and Algorithms

Course Introduction. Algorithm 8/31/2017. COSC 320 Advanced Data Structures and Algorithms. COSC 320 Advanced Data Structures and Algorithms Course Introducton Course Topcs Exams, abs, Proects A quc loo at a few algorthms 1 Advanced Data Structures and Algorthms Descrpton: We are gong to dscuss algorthm complexty analyss, algorthm desgn technques

More information

Support Vector Machines

Support Vector Machines Support Vector Machnes Decson surface s a hyperplane (lne n 2D) n feature space (smlar to the Perceptron) Arguably, the most mportant recent dscovery n machne learnng In a nutshell: map the data to a predetermned

More information

CHAPTER 3 SEQUENTIAL MINIMAL OPTIMIZATION TRAINED SUPPORT VECTOR CLASSIFIER FOR CANCER PREDICTION

CHAPTER 3 SEQUENTIAL MINIMAL OPTIMIZATION TRAINED SUPPORT VECTOR CLASSIFIER FOR CANCER PREDICTION 48 CHAPTER 3 SEQUENTIAL MINIMAL OPTIMIZATION TRAINED SUPPORT VECTOR CLASSIFIER FOR CANCER PREDICTION 3.1 INTRODUCTION The raw mcroarray data s bascally an mage wth dfferent colors ndcatng hybrdzaton (Xue

More information

Classification / Regression Support Vector Machines

Classification / Regression Support Vector Machines Classfcaton / Regresson Support Vector Machnes Jeff Howbert Introducton to Machne Learnng Wnter 04 Topcs SVM classfers for lnearly separable classes SVM classfers for non-lnearly separable classes SVM

More information

Outline. Discriminative classifiers for image recognition. Where in the World? A nearest neighbor recognition example 4/14/2011. CS 376 Lecture 22 1

Outline. Discriminative classifiers for image recognition. Where in the World? A nearest neighbor recognition example 4/14/2011. CS 376 Lecture 22 1 4/14/011 Outlne Dscrmnatve classfers for mage recognton Wednesday, Aprl 13 Krsten Grauman UT-Austn Last tme: wndow-based generc obect detecton basc ppelne face detecton wth boostng as case study Today:

More information

A Binarization Algorithm specialized on Document Images and Photos

A Binarization Algorithm specialized on Document Images and Photos A Bnarzaton Algorthm specalzed on Document mages and Photos Ergna Kavalleratou Dept. of nformaton and Communcaton Systems Engneerng Unversty of the Aegean kavalleratou@aegean.gr Abstract n ths paper, a

More information

Solving two-person zero-sum game by Matlab

Solving two-person zero-sum game by Matlab Appled Mechancs and Materals Onlne: 2011-02-02 ISSN: 1662-7482, Vols. 50-51, pp 262-265 do:10.4028/www.scentfc.net/amm.50-51.262 2011 Trans Tech Publcatons, Swtzerland Solvng two-person zero-sum game by

More information

Abstract Ths paper ponts out an mportant source of necency n Smola and Scholkopf's Sequental Mnmal Optmzaton (SMO) algorthm for SVM regresson that s c

Abstract Ths paper ponts out an mportant source of necency n Smola and Scholkopf's Sequental Mnmal Optmzaton (SMO) algorthm for SVM regresson that s c Improvements to SMO Algorthm for SVM Regresson 1 S.K. Shevade S.S. Keerth C. Bhattacharyya & K.R.K. Murthy shrsh@csa.sc.ernet.n mpessk@guppy.mpe.nus.edu.sg cbchru@csa.sc.ernet.n murthy@csa.sc.ernet.n 1

More information

An Application of the Dulmage-Mendelsohn Decomposition to Sparse Null Space Bases of Full Row Rank Matrices

An Application of the Dulmage-Mendelsohn Decomposition to Sparse Null Space Bases of Full Row Rank Matrices Internatonal Mathematcal Forum, Vol 7, 2012, no 52, 2549-2554 An Applcaton of the Dulmage-Mendelsohn Decomposton to Sparse Null Space Bases of Full Row Rank Matrces Mostafa Khorramzadeh Department of Mathematcal

More information

Problem Definitions and Evaluation Criteria for Computational Expensive Optimization

Problem Definitions and Evaluation Criteria for Computational Expensive Optimization Problem efntons and Evaluaton Crtera for Computatonal Expensve Optmzaton B. Lu 1, Q. Chen and Q. Zhang 3, J. J. Lang 4, P. N. Suganthan, B. Y. Qu 6 1 epartment of Computng, Glyndwr Unversty, UK Faclty

More information

Edge Detection in Noisy Images Using the Support Vector Machines

Edge Detection in Noisy Images Using the Support Vector Machines Edge Detecton n Nosy Images Usng the Support Vector Machnes Hlaro Gómez-Moreno, Saturnno Maldonado-Bascón, Francsco López-Ferreras Sgnal Theory and Communcatons Department. Unversty of Alcalá Crta. Madrd-Barcelona

More information

The Research of Support Vector Machine in Agricultural Data Classification

The Research of Support Vector Machine in Agricultural Data Classification The Research of Support Vector Machne n Agrcultural Data Classfcaton Le Sh, Qguo Duan, Xnmng Ma, Me Weng College of Informaton and Management Scence, HeNan Agrcultural Unversty, Zhengzhou 45000 Chna Zhengzhou

More information

Parallel matrix-vector multiplication

Parallel matrix-vector multiplication Appendx A Parallel matrx-vector multplcaton The reduced transton matrx of the three-dmensonal cage model for gel electrophoress, descrbed n secton 3.2, becomes excessvely large for polymer lengths more

More information

Classifier Selection Based on Data Complexity Measures *

Classifier Selection Based on Data Complexity Measures * Classfer Selecton Based on Data Complexty Measures * Edth Hernández-Reyes, J.A. Carrasco-Ochoa, and J.Fco. Martínez-Trndad Natonal Insttute for Astrophyscs, Optcs and Electroncs, Lus Enrque Erro No.1 Sta.

More information

Machine Learning. Support Vector Machines. (contains material adapted from talks by Constantin F. Aliferis & Ioannis Tsamardinos, and Martin Law)

Machine Learning. Support Vector Machines. (contains material adapted from talks by Constantin F. Aliferis & Ioannis Tsamardinos, and Martin Law) Machne Learnng Support Vector Machnes (contans materal adapted from talks by Constantn F. Alfers & Ioanns Tsamardnos, and Martn Law) Bryan Pardo, Machne Learnng: EECS 349 Fall 2014 Support Vector Machnes

More information

Classifying Acoustic Transient Signals Using Artificial Intelligence

Classifying Acoustic Transient Signals Using Artificial Intelligence Classfyng Acoustc Transent Sgnals Usng Artfcal Intellgence Steve Sutton, Unversty of North Carolna At Wlmngton (suttons@charter.net) Greg Huff, Unversty of North Carolna At Wlmngton (jgh7476@uncwl.edu)

More information

Skew Angle Estimation and Correction of Hand Written, Textual and Large areas of Non-Textual Document Images: A Novel Approach

Skew Angle Estimation and Correction of Hand Written, Textual and Large areas of Non-Textual Document Images: A Novel Approach Angle Estmaton and Correcton of Hand Wrtten, Textual and Large areas of Non-Textual Document Images: A Novel Approach D.R.Ramesh Babu Pyush M Kumat Mahesh D Dhannawat PES Insttute of Technology Research

More information

Lecture 5: Multilayer Perceptrons

Lecture 5: Multilayer Perceptrons Lecture 5: Multlayer Perceptrons Roger Grosse 1 Introducton So far, we ve only talked about lnear models: lnear regresson and lnear bnary classfers. We noted that there are functons that can t be represented

More information

BOOSTING CLASSIFICATION ACCURACY WITH SAMPLES CHOSEN FROM A VALIDATION SET

BOOSTING CLASSIFICATION ACCURACY WITH SAMPLES CHOSEN FROM A VALIDATION SET 1 BOOSTING CLASSIFICATION ACCURACY WITH SAMPLES CHOSEN FROM A VALIDATION SET TZU-CHENG CHUANG School of Electrcal and Computer Engneerng, Purdue Unversty, West Lafayette, Indana 47907 SAUL B. GELFAND School

More information

Programming in Fortran 90 : 2017/2018

Programming in Fortran 90 : 2017/2018 Programmng n Fortran 90 : 2017/2018 Programmng n Fortran 90 : 2017/2018 Exercse 1 : Evaluaton of functon dependng on nput Wrte a program who evaluate the functon f (x,y) for any two user specfed values

More information

Smoothing Spline ANOVA for variable screening

Smoothing Spline ANOVA for variable screening Smoothng Splne ANOVA for varable screenng a useful tool for metamodels tranng and mult-objectve optmzaton L. Rcco, E. Rgon, A. Turco Outlne RSM Introducton Possble couplng Test case MOO MOO wth Game Theory

More information

NGPM -- A NSGA-II Program in Matlab

NGPM -- A NSGA-II Program in Matlab Verson 1.4 LIN Song Aerospace Structural Dynamcs Research Laboratory College of Astronautcs, Northwestern Polytechncal Unversty, Chna Emal: lsssswc@163.com 2011-07-26 Contents Contents... 1. Introducton...

More information

Three supervised learning methods on pen digits character recognition dataset

Three supervised learning methods on pen digits character recognition dataset Three supervsed learnng methods on pen dgts character recognton dataset Chrs Flezach Department of Computer Scence and Engneerng Unversty of Calforna, San Dego San Dego, CA 92093 cflezac@cs.ucsd.edu Satoru

More information

Concurrent Apriori Data Mining Algorithms

Concurrent Apriori Data Mining Algorithms Concurrent Apror Data Mnng Algorthms Vassl Halatchev Department of Electrcal Engneerng and Computer Scence York Unversty, Toronto October 8, 2015 Outlne Why t s mportant Introducton to Assocaton Rule Mnng

More information

Cluster Analysis of Electrical Behavior

Cluster Analysis of Electrical Behavior Journal of Computer and Communcatons, 205, 3, 88-93 Publshed Onlne May 205 n ScRes. http://www.scrp.org/ournal/cc http://dx.do.org/0.4236/cc.205.350 Cluster Analyss of Electrcal Behavor Ln Lu Ln Lu, School

More information

Announcements. Supervised Learning

Announcements. Supervised Learning Announcements See Chapter 5 of Duda, Hart, and Stork. Tutoral by Burge lnked to on web page. Supervsed Learnng Classfcaton wth labeled eamples. Images vectors n hgh-d space. Supervsed Learnng Labeled eamples

More information

CMPS 10 Introduction to Computer Science Lecture Notes

CMPS 10 Introduction to Computer Science Lecture Notes CPS 0 Introducton to Computer Scence Lecture Notes Chapter : Algorthm Desgn How should we present algorthms? Natural languages lke Englsh, Spansh, or French whch are rch n nterpretaton and meanng are not

More information

Term Weighting Classification System Using the Chi-square Statistic for the Classification Subtask at NTCIR-6 Patent Retrieval Task

Term Weighting Classification System Using the Chi-square Statistic for the Classification Subtask at NTCIR-6 Patent Retrieval Task Proceedngs of NTCIR-6 Workshop Meetng, May 15-18, 2007, Tokyo, Japan Term Weghtng Classfcaton System Usng the Ch-square Statstc for the Classfcaton Subtask at NTCIR-6 Patent Retreval Task Kotaro Hashmoto

More information

Discriminative Dictionary Learning with Pairwise Constraints

Discriminative Dictionary Learning with Pairwise Constraints Dscrmnatve Dctonary Learnng wth Parwse Constrants Humn Guo Zhuoln Jang LARRY S. DAVIS UNIVERSITY OF MARYLAND Nov. 6 th, Outlne Introducton/motvaton Dctonary Learnng Dscrmnatve Dctonary Learnng wth Parwse

More information

Sum of Linear and Fractional Multiobjective Programming Problem under Fuzzy Rules Constraints

Sum of Linear and Fractional Multiobjective Programming Problem under Fuzzy Rules Constraints Australan Journal of Basc and Appled Scences, 2(4): 1204-1208, 2008 ISSN 1991-8178 Sum of Lnear and Fractonal Multobjectve Programmng Problem under Fuzzy Rules Constrants 1 2 Sanjay Jan and Kalash Lachhwan

More information

The Greedy Method. Outline and Reading. Change Money Problem. Greedy Algorithms. Applications of the Greedy Strategy. The Greedy Method Technique

The Greedy Method. Outline and Reading. Change Money Problem. Greedy Algorithms. Applications of the Greedy Strategy. The Greedy Method Technique //00 :0 AM Outlne and Readng The Greedy Method The Greedy Method Technque (secton.) Fractonal Knapsack Problem (secton..) Task Schedulng (secton..) Mnmum Spannng Trees (secton.) Change Money Problem Greedy

More information

Improvement of Spatial Resolution Using BlockMatching Based Motion Estimation and Frame. Integration

Improvement of Spatial Resolution Using BlockMatching Based Motion Estimation and Frame. Integration Improvement of Spatal Resoluton Usng BlockMatchng Based Moton Estmaton and Frame Integraton Danya Suga and Takayuk Hamamoto Graduate School of Engneerng, Tokyo Unversty of Scence, 6-3-1, Nuku, Katsuska-ku,

More information

Efficient Distributed Linear Classification Algorithms via the Alternating Direction Method of Multipliers

Efficient Distributed Linear Classification Algorithms via the Alternating Direction Method of Multipliers Effcent Dstrbuted Lnear Classfcaton Algorthms va the Alternatng Drecton Method of Multplers Caoxe Zhang Honglak Lee Kang G. Shn Department of EECS Unversty of Mchgan Ann Arbor, MI 48109, USA caoxezh@umch.edu

More information

Efficient Text Classification by Weighted Proximal SVM *

Efficient Text Classification by Weighted Proximal SVM * Effcent ext Classfcaton by Weghted Proxmal SVM * Dong Zhuang 1, Benyu Zhang, Qang Yang 3, Jun Yan 4, Zheng Chen, Yng Chen 1 1 Computer Scence and Engneerng, Bejng Insttute of echnology, Bejng 100081, Chna

More information

A Modified Median Filter for the Removal of Impulse Noise Based on the Support Vector Machines

A Modified Median Filter for the Removal of Impulse Noise Based on the Support Vector Machines A Modfed Medan Flter for the Removal of Impulse Nose Based on the Support Vector Machnes H. GOMEZ-MORENO, S. MALDONADO-BASCON, F. LOPEZ-FERRERAS, M. UTRILLA- MANSO AND P. GIL-JIMENEZ Departamento de Teoría

More information

Network Intrusion Detection Based on PSO-SVM

Network Intrusion Detection Based on PSO-SVM TELKOMNIKA Indonesan Journal of Electrcal Engneerng Vol.1, No., February 014, pp. 150 ~ 1508 DOI: http://dx.do.org/10.11591/telkomnka.v1.386 150 Network Intruson Detecton Based on PSO-SVM Changsheng Xang*

More information

An Optimal Algorithm for Prufer Codes *

An Optimal Algorithm for Prufer Codes * J. Software Engneerng & Applcatons, 2009, 2: 111-115 do:10.4236/jsea.2009.22016 Publshed Onlne July 2009 (www.scrp.org/journal/jsea) An Optmal Algorthm for Prufer Codes * Xaodong Wang 1, 2, Le Wang 3,

More information

A fault tree analysis strategy using binary decision diagrams

A fault tree analysis strategy using binary decision diagrams Loughborough Unversty Insttutonal Repostory A fault tree analyss strategy usng bnary decson dagrams Ths tem was submtted to Loughborough Unversty's Insttutonal Repostory by the/an author. Addtonal Informaton:

More information

Data Mining: Model Evaluation

Data Mining: Model Evaluation Data Mnng: Model Evaluaton Aprl 16, 2013 1 Issues: Evaluatng Classfcaton Methods Accurac classfer accurac: predctng class label predctor accurac: guessng value of predcted attrbutes Speed tme to construct

More information

Using Neural Networks and Support Vector Machines in Data Mining

Using Neural Networks and Support Vector Machines in Data Mining Usng eural etworks and Support Vector Machnes n Data Mnng RICHARD A. WASIOWSKI Computer Scence Department Calforna State Unversty Domnguez Hlls Carson, CA 90747 USA Abstract: - Multvarate data analyss

More information

Module Management Tool in Software Development Organizations

Module Management Tool in Software Development Organizations Journal of Computer Scence (5): 8-, 7 ISSN 59-66 7 Scence Publcatons Management Tool n Software Development Organzatons Ahmad A. Al-Rababah and Mohammad A. Al-Rababah Faculty of IT, Al-Ahlyyah Amman Unversty,

More information

Optimizing Document Scoring for Query Retrieval

Optimizing Document Scoring for Query Retrieval Optmzng Document Scorng for Query Retreval Brent Ellwen baellwe@cs.stanford.edu Abstract The goal of ths project was to automate the process of tunng a document query engne. Specfcally, I used machne learnng

More information

Wavefront Reconstructor

Wavefront Reconstructor A Dstrbuted Smplex B-Splne Based Wavefront Reconstructor Coen de Vsser and Mchel Verhaegen 14-12-201212 2012 Delft Unversty of Technology Contents Introducton Wavefront reconstructon usng Smplex B-Splnes

More information

CS246: Mining Massive Datasets Jure Leskovec, Stanford University

CS246: Mining Massive Datasets Jure Leskovec, Stanford University CS46: Mnng Massve Datasets Jure Leskovec, Stanford Unversty http://cs46.stanford.edu /19/013 Jure Leskovec, Stanford CS46: Mnng Massve Datasets, http://cs46.stanford.edu Perceptron: y = sgn( x Ho to fnd

More information

A Unified Framework for Semantics and Feature Based Relevance Feedback in Image Retrieval Systems

A Unified Framework for Semantics and Feature Based Relevance Feedback in Image Retrieval Systems A Unfed Framework for Semantcs and Feature Based Relevance Feedback n Image Retreval Systems Ye Lu *, Chunhu Hu 2, Xngquan Zhu 3*, HongJang Zhang 2, Qang Yang * School of Computng Scence Smon Fraser Unversty

More information

Incremental Learning with Support Vector Machines and Fuzzy Set Theory

Incremental Learning with Support Vector Machines and Fuzzy Set Theory The 25th Workshop on Combnatoral Mathematcs and Computaton Theory Incremental Learnng wth Support Vector Machnes and Fuzzy Set Theory Yu-Mng Chuang 1 and Cha-Hwa Ln 2* 1 Department of Computer Scence and

More information

An Efficient Genetic Algorithm with Fuzzy c-means Clustering for Traveling Salesman Problem

An Efficient Genetic Algorithm with Fuzzy c-means Clustering for Traveling Salesman Problem An Effcent Genetc Algorthm wth Fuzzy c-means Clusterng for Travelng Salesman Problem Jong-Won Yoon and Sung-Bae Cho Dept. of Computer Scence Yonse Unversty Seoul, Korea jwyoon@sclab.yonse.ac.r, sbcho@cs.yonse.ac.r

More information

Assembler. Building a Modern Computer From First Principles.

Assembler. Building a Modern Computer From First Principles. Assembler Buldng a Modern Computer From Frst Prncples www.nand2tetrs.org Elements of Computng Systems, Nsan & Schocken, MIT Press, www.nand2tetrs.org, Chapter 6: Assembler slde Where we are at: Human Thought

More information

Outline. Type of Machine Learning. Examples of Application. Unsupervised Learning

Outline. Type of Machine Learning. Examples of Application. Unsupervised Learning Outlne Artfcal Intellgence and ts applcatons Lecture 8 Unsupervsed Learnng Professor Danel Yeung danyeung@eee.org Dr. Patrck Chan patrckchan@eee.org South Chna Unversty of Technology, Chna Introducton

More information

Determining the Optimal Bandwidth Based on Multi-criterion Fusion

Determining the Optimal Bandwidth Based on Multi-criterion Fusion Proceedngs of 01 4th Internatonal Conference on Machne Learnng and Computng IPCSIT vol. 5 (01) (01) IACSIT Press, Sngapore Determnng the Optmal Bandwdth Based on Mult-crteron Fuson Ha-L Lang 1+, Xan-Mn

More information

Data Mining For Multi-Criteria Energy Predictions

Data Mining For Multi-Criteria Energy Predictions Data Mnng For Mult-Crtera Energy Predctons Kashf Gll and Denns Moon Abstract We present a data mnng technque for mult-crtera predctons of wnd energy. A mult-crtera (MC) evolutonary computng method has

More information

SLAM Summer School 2006 Practical 2: SLAM using Monocular Vision

SLAM Summer School 2006 Practical 2: SLAM using Monocular Vision SLAM Summer School 2006 Practcal 2: SLAM usng Monocular Vson Javer Cvera, Unversty of Zaragoza Andrew J. Davson, Imperal College London J.M.M Montel, Unversty of Zaragoza. josemar@unzar.es, jcvera@unzar.es,

More information

GSLM Operations Research II Fall 13/14

GSLM Operations Research II Fall 13/14 GSLM 58 Operatons Research II Fall /4 6. Separable Programmng Consder a general NLP mn f(x) s.t. g j (x) b j j =. m. Defnton 6.. The NLP s a separable program f ts objectve functon and all constrants are

More information

Feature Selection as an Improving Step for Decision Tree Construction

Feature Selection as an Improving Step for Decision Tree Construction 2009 Internatonal Conference on Machne Learnng and Computng IPCSIT vol.3 (2011) (2011) IACSIT Press, Sngapore Feature Selecton as an Improvng Step for Decson Tree Constructon Mahd Esmael 1, Fazekas Gabor

More information

Assignment # 2. Farrukh Jabeen Algorithms 510 Assignment #2 Due Date: June 15, 2009.

Assignment # 2. Farrukh Jabeen Algorithms 510 Assignment #2 Due Date: June 15, 2009. Farrukh Jabeen Algorthms 51 Assgnment #2 Due Date: June 15, 29. Assgnment # 2 Chapter 3 Dscrete Fourer Transforms Implement the FFT for the DFT. Descrbed n sectons 3.1 and 3.2. Delverables: 1. Concse descrpton

More information

The Codesign Challenge

The Codesign Challenge ECE 4530 Codesgn Challenge Fall 2007 Hardware/Software Codesgn The Codesgn Challenge Objectves In the codesgn challenge, your task s to accelerate a gven software reference mplementaton as fast as possble.

More information

The Shortest Path of Touring Lines given in the Plane

The Shortest Path of Touring Lines given in the Plane Send Orders for Reprnts to reprnts@benthamscence.ae 262 The Open Cybernetcs & Systemcs Journal, 2015, 9, 262-267 The Shortest Path of Tourng Lnes gven n the Plane Open Access Ljuan Wang 1,2, Dandan He

More information

Journal of Chemical and Pharmaceutical Research, 2014, 6(6): Research Article. A selective ensemble classification method on microarray data

Journal of Chemical and Pharmaceutical Research, 2014, 6(6): Research Article. A selective ensemble classification method on microarray data Avalable onlne www.ocpr.com Journal of Chemcal and Pharmaceutcal Research, 2014, 6(6):2860-2866 Research Artcle ISSN : 0975-7384 CODEN(USA) : JCPRC5 A selectve ensemble classfcaton method on mcroarray

More information

Sequential search. Building Java Programs Chapter 13. Sequential search. Sequential search

Sequential search. Building Java Programs Chapter 13. Sequential search. Sequential search Sequental search Buldng Java Programs Chapter 13 Searchng and Sortng sequental search: Locates a target value n an array/lst by examnng each element from start to fnsh. How many elements wll t need to

More information

Compiler Design. Spring Register Allocation. Sample Exercises and Solutions. Prof. Pedro C. Diniz

Compiler Design. Spring Register Allocation. Sample Exercises and Solutions. Prof. Pedro C. Diniz Compler Desgn Sprng 2014 Regster Allocaton Sample Exercses and Solutons Prof. Pedro C. Dnz USC / Informaton Scences Insttute 4676 Admralty Way, Sute 1001 Marna del Rey, Calforna 90292 pedro@s.edu Regster

More information

From Comparing Clusterings to Combining Clusterings

From Comparing Clusterings to Combining Clusterings Proceedngs of the Twenty-Thrd AAAI Conference on Artfcal Intellgence (008 From Comparng Clusterngs to Combnng Clusterngs Zhwu Lu and Yuxn Peng and Janguo Xao Insttute of Computer Scence and Technology,

More information

S1 Note. Basis functions.

S1 Note. Basis functions. S1 Note. Bass functons. Contents Types of bass functons...1 The Fourer bass...2 B-splne bass...3 Power and type I error rates wth dfferent numbers of bass functons...4 Table S1. Smulaton results of type

More information

Steps for Computing the Dissimilarity, Entropy, Herfindahl-Hirschman and. Accessibility (Gravity with Competition) Indices

Steps for Computing the Dissimilarity, Entropy, Herfindahl-Hirschman and. Accessibility (Gravity with Competition) Indices Steps for Computng the Dssmlarty, Entropy, Herfndahl-Hrschman and Accessblty (Gravty wth Competton) Indces I. Dssmlarty Index Measurement: The followng formula can be used to measure the evenness between

More information

Alternating Direction Method of Multipliers Implementation Using Apache Spark

Alternating Direction Method of Multipliers Implementation Using Apache Spark Alternatng Drecton Method of Multplers Implementaton Usng Apache Spark Deterch Lawson June 4, 2014 1 Introducton Many applcaton areas n optmzaton have benefted from recent trends towards massve datasets.

More information

Learning a Class-Specific Dictionary for Facial Expression Recognition

Learning a Class-Specific Dictionary for Facial Expression Recognition BULGARIAN ACADEMY OF SCIENCES CYBERNETICS AND INFORMATION TECHNOLOGIES Volume 16, No 4 Sofa 016 Prnt ISSN: 1311-970; Onlne ISSN: 1314-4081 DOI: 10.1515/cat-016-0067 Learnng a Class-Specfc Dctonary for

More information

Tsinghua University at TAC 2009: Summarizing Multi-documents by Information Distance

Tsinghua University at TAC 2009: Summarizing Multi-documents by Information Distance Tsnghua Unversty at TAC 2009: Summarzng Mult-documents by Informaton Dstance Chong Long, Mnle Huang, Xaoyan Zhu State Key Laboratory of Intellgent Technology and Systems, Tsnghua Natonal Laboratory for

More information

Load-Balanced Anycast Routing

Load-Balanced Anycast Routing Load-Balanced Anycast Routng Chng-Yu Ln, Jung-Hua Lo, and Sy-Yen Kuo Department of Electrcal Engneerng atonal Tawan Unversty, Tape, Tawan sykuo@cc.ee.ntu.edu.tw Abstract For fault-tolerance and load-balance

More information

Subspace clustering. Clustering. Fundamental to all clustering techniques is the choice of distance measure between data points;

Subspace clustering. Clustering. Fundamental to all clustering techniques is the choice of distance measure between data points; Subspace clusterng Clusterng Fundamental to all clusterng technques s the choce of dstance measure between data ponts; D q ( ) ( ) 2 x x = x x, j k = 1 k jk Squared Eucldean dstance Assumpton: All features

More information

A Fast Content-Based Multimedia Retrieval Technique Using Compressed Data

A Fast Content-Based Multimedia Retrieval Technique Using Compressed Data A Fast Content-Based Multmeda Retreval Technque Usng Compressed Data Borko Furht and Pornvt Saksobhavvat NSF Multmeda Laboratory Florda Atlantc Unversty, Boca Raton, Florda 3343 ABSTRACT In ths paper,

More information

Array transposition in CUDA shared memory

Array transposition in CUDA shared memory Array transposton n CUDA shared memory Mke Gles February 19, 2014 Abstract Ths short note s nspred by some code wrtten by Jeremy Appleyard for the transposton of data through shared memory. I had some

More information

Random Kernel Perceptron on ATTiny2313 Microcontroller

Random Kernel Perceptron on ATTiny2313 Microcontroller Random Kernel Perceptron on ATTny233 Mcrocontroller Nemanja Djurc Department of Computer and Informaton Scences, Temple Unversty Phladelpha, PA 922, USA nemanja.djurc@temple.edu Slobodan Vucetc Department

More information

Exercises (Part 4) Introduction to R UCLA/CCPR. John Fox, February 2005

Exercises (Part 4) Introduction to R UCLA/CCPR. John Fox, February 2005 Exercses (Part 4) Introducton to R UCLA/CCPR John Fox, February 2005 1. A challengng problem: Iterated weghted least squares (IWLS) s a standard method of fttng generalzed lnear models to data. As descrbed

More information

DECISION SUPPORT SYSTEM FOR HEART DISEASE BASED ON SEQUENTIAL MINIMAL OPTIMIZATION IN SUPPORT VECTOR MACHINE

DECISION SUPPORT SYSTEM FOR HEART DISEASE BASED ON SEQUENTIAL MINIMAL OPTIMIZATION IN SUPPORT VECTOR MACHINE DECISION SUPPORT SYSTEM FOR HEART DISEASE BASED ON SEQUENTIAL MINIMAL OPTIMIZATION IN SUPPORT VECTOR MACHINE Deept Vadcherla, Sheetal Sonawane Department of Computer Engneerng, Pune Insttute of Computer

More information

Fuzzy Modeling of the Complexity vs. Accuracy Trade-off in a Sequential Two-Stage Multi-Classifier System

Fuzzy Modeling of the Complexity vs. Accuracy Trade-off in a Sequential Two-Stage Multi-Classifier System Fuzzy Modelng of the Complexty vs. Accuracy Trade-off n a Sequental Two-Stage Mult-Classfer System MARK LAST 1 Department of Informaton Systems Engneerng Ben-Guron Unversty of the Negev Beer-Sheva 84105

More information

Feature Reduction and Selection

Feature Reduction and Selection Feature Reducton and Selecton Dr. Shuang LIANG School of Software Engneerng TongJ Unversty Fall, 2012 Today s Topcs Introducton Problems of Dmensonalty Feature Reducton Statstc methods Prncpal Components

More information

5 The Primal-Dual Method

5 The Primal-Dual Method 5 The Prmal-Dual Method Orgnally desgned as a method for solvng lnear programs, where t reduces weghted optmzaton problems to smpler combnatoral ones, the prmal-dual method (PDM) has receved much attenton

More information

Parallel Numerics. 1 Preconditioning & Iterative Solvers (From 2016)

Parallel Numerics. 1 Preconditioning & Iterative Solvers (From 2016) Technsche Unverstät München WSe 6/7 Insttut für Informatk Prof. Dr. Thomas Huckle Dpl.-Math. Benjamn Uekermann Parallel Numercs Exercse : Prevous Exam Questons Precondtonng & Iteratve Solvers (From 6)

More information

A mathematical programming approach to the analysis, design and scheduling of offshore oilfields

A mathematical programming approach to the analysis, design and scheduling of offshore oilfields 17 th European Symposum on Computer Aded Process Engneerng ESCAPE17 V. Plesu and P.S. Agach (Edtors) 2007 Elsever B.V. All rghts reserved. 1 A mathematcal programmng approach to the analyss, desgn and

More information

BAYESIAN MULTI-SOURCE DOMAIN ADAPTATION

BAYESIAN MULTI-SOURCE DOMAIN ADAPTATION BAYESIAN MULTI-SOURCE DOMAIN ADAPTATION SHI-LIANG SUN, HONG-LEI SHI Department of Computer Scence and Technology, East Chna Normal Unversty 500 Dongchuan Road, Shangha 200241, P. R. Chna E-MAIL: slsun@cs.ecnu.edu.cn,

More information

CHAPTER 2 PROPOSED IMPROVED PARTICLE SWARM OPTIMIZATION

CHAPTER 2 PROPOSED IMPROVED PARTICLE SWARM OPTIMIZATION 24 CHAPTER 2 PROPOSED IMPROVED PARTICLE SWARM OPTIMIZATION The present chapter proposes an IPSO approach for multprocessor task schedulng problem wth two classfcatons, namely, statc ndependent tasks and

More information

Journal of Process Control

Journal of Process Control Journal of Process Control (0) 738 750 Contents lsts avalable at ScVerse ScenceDrect Journal of Process Control j ourna l ho me pag e: wwwelsevercom/locate/jprocont Decentralzed fault detecton and dagnoss

More information

Comparison of Heuristics for Scheduling Independent Tasks on Heterogeneous Distributed Environments

Comparison of Heuristics for Scheduling Independent Tasks on Heterogeneous Distributed Environments Comparson of Heurstcs for Schedulng Independent Tasks on Heterogeneous Dstrbuted Envronments Hesam Izakan¹, Ath Abraham², Senor Member, IEEE, Václav Snášel³ ¹ Islamc Azad Unversty, Ramsar Branch, Ramsar,

More information

ISSN: International Journal of Engineering and Innovative Technology (IJEIT) Volume 1, Issue 4, April 2012

ISSN: International Journal of Engineering and Innovative Technology (IJEIT) Volume 1, Issue 4, April 2012 Performance Evoluton of Dfferent Codng Methods wth β - densty Decodng Usng Error Correctng Output Code Based on Multclass Classfcaton Devangn Dave, M. Samvatsar, P. K. Bhanoda Abstract A common way to

More information

Fast Computation of Shortest Path for Visiting Segments in the Plane

Fast Computation of Shortest Path for Visiting Segments in the Plane Send Orders for Reprnts to reprnts@benthamscence.ae 4 The Open Cybernetcs & Systemcs Journal, 04, 8, 4-9 Open Access Fast Computaton of Shortest Path for Vstng Segments n the Plane Ljuan Wang,, Bo Jang

More information

Multi-objective Optimization Using Self-adaptive Differential Evolution Algorithm

Multi-objective Optimization Using Self-adaptive Differential Evolution Algorithm Mult-objectve Optmzaton Usng Self-adaptve Dfferental Evoluton Algorthm V. L. Huang, S. Z. Zhao, R. Mallpedd and P. N. Suganthan Abstract - In ths paper, we propose a Multobjectve Self-adaptve Dfferental

More information

Empirical Distributions of Parameter Estimates. in Binary Logistic Regression Using Bootstrap

Empirical Distributions of Parameter Estimates. in Binary Logistic Regression Using Bootstrap Int. Journal of Math. Analyss, Vol. 8, 4, no. 5, 7-7 HIKARI Ltd, www.m-hkar.com http://dx.do.org/.988/jma.4.494 Emprcal Dstrbutons of Parameter Estmates n Bnary Logstc Regresson Usng Bootstrap Anwar Ftranto*

More information

Learning to Project in Multi-Objective Binary Linear Programming

Learning to Project in Multi-Objective Binary Linear Programming Learnng to Project n Mult-Objectve Bnary Lnear Programmng Alvaro Serra-Altamranda Department of Industral and Management System Engneerng, Unversty of South Florda, Tampa, FL, 33620 USA, amserra@mal.usf.edu,

More information

Kent State University CS 4/ Design and Analysis of Algorithms. Dept. of Math & Computer Science LECT-16. Dynamic Programming

Kent State University CS 4/ Design and Analysis of Algorithms. Dept. of Math & Computer Science LECT-16. Dynamic Programming CS 4/560 Desgn and Analyss of Algorthms Kent State Unversty Dept. of Math & Computer Scence LECT-6 Dynamc Programmng 2 Dynamc Programmng Dynamc Programmng, lke the dvde-and-conquer method, solves problems

More information

A fast algorithm for color image segmentation

A fast algorithm for color image segmentation Unersty of Wollongong Research Onlne Faculty of Informatcs - Papers (Arche) Faculty of Engneerng and Informaton Scences 006 A fast algorthm for color mage segmentaton L. Dong Unersty of Wollongong, lju@uow.edu.au

More information

CSCI 104 Sorting Algorithms. Mark Redekopp David Kempe

CSCI 104 Sorting Algorithms. Mark Redekopp David Kempe CSCI 104 Sortng Algorthms Mark Redekopp Davd Kempe Algorthm Effcency SORTING 2 Sortng If we have an unordered lst, sequental search becomes our only choce If we wll perform a lot of searches t may be benefcal

More information

Discriminative classifiers for object classification. Last time

Discriminative classifiers for object classification. Last time Dscrmnatve classfers for object classfcaton Thursday, Nov 12 Krsten Grauman UT Austn Last tme Supervsed classfcaton Loss and rsk, kbayes rule Skn color detecton example Sldng ndo detecton Classfers, boostng

More information

Complex System Reliability Evaluation using Support Vector Machine for Incomplete Data-set

Complex System Reliability Evaluation using Support Vector Machine for Incomplete Data-set Internatonal Journal of Performablty Engneerng, Vol. 7, No. 1, January 2010, pp.32-42. RAMS Consultants Prnted n Inda Complex System Relablty Evaluaton usng Support Vector Machne for Incomplete Data-set

More information

Human Face Recognition Using Generalized. Kernel Fisher Discriminant

Human Face Recognition Using Generalized. Kernel Fisher Discriminant Human Face Recognton Usng Generalzed Kernel Fsher Dscrmnant ng-yu Sun,2 De-Shuang Huang Ln Guo. Insttute of Intellgent Machnes, Chnese Academy of Scences, P.O.ox 30, Hefe, Anhu, Chna. 2. Department of

More information

NUMERICAL SOLVING OPTIMAL CONTROL PROBLEMS BY THE METHOD OF VARIATIONS

NUMERICAL SOLVING OPTIMAL CONTROL PROBLEMS BY THE METHOD OF VARIATIONS ARPN Journal of Engneerng and Appled Scences 006-017 Asan Research Publshng Network (ARPN). All rghts reserved. NUMERICAL SOLVING OPTIMAL CONTROL PROBLEMS BY THE METHOD OF VARIATIONS Igor Grgoryev, Svetlana

More information

Content Based Image Retrieval Using 2-D Discrete Wavelet with Texture Feature with Different Classifiers

Content Based Image Retrieval Using 2-D Discrete Wavelet with Texture Feature with Different Classifiers IOSR Journal of Electroncs and Communcaton Engneerng (IOSR-JECE) e-issn: 78-834,p- ISSN: 78-8735.Volume 9, Issue, Ver. IV (Mar - Apr. 04), PP 0-07 Content Based Image Retreval Usng -D Dscrete Wavelet wth

More information

Virtual Machine Migration based on Trust Measurement of Computer Node

Virtual Machine Migration based on Trust Measurement of Computer Node Appled Mechancs and Materals Onlne: 2014-04-04 ISSN: 1662-7482, Vols. 536-537, pp 678-682 do:10.4028/www.scentfc.net/amm.536-537.678 2014 Trans Tech Publcatons, Swtzerland Vrtual Machne Mgraton based on

More information