Large-scale Distance Metric Learning with Uncertainty

Size: px
Start display at page:

Download "Large-scale Distance Metric Learning with Uncertainty"

Transcription

1 i Large-scale Disance Meric Learning wih Uncerainy Qi Qian Jiasheng Tang Ha Li Shenghu Zhu Rng Jin Alibaba Grup, Bellevue, WA, 98004, USA {qi.qian, jiasheng.js, liha.lh, shenghu.zhu, Opimizing he meric wih a se f riple cnsrains is challenging since he number f riple cnsrains can be up O(n 3 ), where n is he number f he riginal raining examples. I makes DML cmpuainally inracable fr he large-scale prblems. Many sraegies have been develped deal wih his challenge and ms f hem fall in w caegries, learning by schasic gradien descen (SGD) and learning wih he acive se. Wih he sraegy f SGD, DML mehds can sample jus ne cnsrain r a mini-bach f cnsrains a each ierain bserve an unbiased esimain f he full gradien and avid cmpuarxiv: v1 [cs.lg] 25 May 2018 Absrac Disance meric learning (DML) has been sudied exensively in he pas decades fr is superir perfrmance wih disance-based algrihms. Ms f he exising mehds prpse learn a disance meric wih pairwise r riple cnsrains. Hwever, he number f cnsrains is quadraic r even cubic in he number f he riginal examples, which makes i challenging fr DML handle he large-scale daa se. Besides, he real-wrld daa may cnain varius uncerainy, especially fr he image daa. The uncerainy can mislead he learning prcedure and cause he perfrmance degradain. By invesigaing he image daa, we find ha he riginal daa can be bserved frm a small se f clean laen examples wih differen disrins. In his wrk, we prpse he margin preserving meric learning framewrk learn he disance meric and laen examples simulaneusly. By leveraging he ideal prperies f laen examples, he raining efficiency can be imprved significanly while he learned meric als becmes rbus he uncerainy in he riginal daa. Furhermre, we can shw ha he meric is learned frm laen examples nly, bu i can preserve he large margin prpery even fr he riginal daa. The empirical sudy n he benchmark image daa ses demnsraes he efficacy and efficiency f he prpsed mehd. 1. Inrducin Disance meric learning (DML) aims learn a disance meric where examples frm he same class are well separaed frm examples f differen classes. I is an essenial ask fr disance-based algrihms, such as k-means clusering [18], k-neares neighbr classificain [17] and infrmain rerieval [2]. Given a disance meric M, he squared Mahalanbis disance beween examples x i and x j can be cmpued as D 2 M (x i, x j ) = (x i x j ) M(x i x j ) Ms f exising DML mehds prpse learn he meric by minimizing he number f vilains in he se f xjx xi xk Figure 1. Illusrain f he prpsed mehd. Le rund and square pins dene he arge daa and impsrs, respecively. Le riangle pins dene he crrespnding laen examples. Daa pins wih he same clr are frm he same class. I demnsraes ha he meric learned wih laen examples n nly separaes he dissimilar laen daa wih a large margin bu als preserves he large margin fr he riginal daa. pairwise r riple cnsrains. Given a se f pairwise cnsrains, DML ries learn a meric such ha he disances beween examples frm he same class are sufficienly small (e.g., smaller han a predefined hreshld) while hse beween differen nes are large enugh [3, 18]. Differen frm pairwise cnsrains, each riple cnsrain cnsiss f hree examples (x i, x j, x k ), where x i and x j have he same label and x k is frm a differen class. An ideal meric can push away x k frm x i and x j by a large margin [17]. Learning wih riple cnsrains pimizes he lcal psiins f examples and is mre flexible fr real-wrld applicains, where defining he apprpriae hreshlds is hard fr pairwise cnsrains. In his wrk, we will fcus n DML wih riple cnsrains. xj x xi xk

2 ing he gradien frm he whle se [2, 10]. Oher mehds learn he meric wih a se f acive cnsrains (i.e., vilaed by he curren meric), where he size can be significanly smaller han he riginal se [17]. I is a cnveninal sraegy applied by cuing plane mehds [1]. Bh f hese sraegies can alleviae he large-scale challenge bu have inheren drawbacks. Appraches based n SGD have search hrugh he whle se f riple cnsrains, which resuls in he slw cnvergence, especially when he number f acive cnsrains is small. On he her hand, he mehds relying n he acive se have idenify he se a each ierain. Unfrunaely, his perain requires cmpuing pairwise disances wih he curren meric, where he cs is O(n 2 ) and is expensive fr large-scale prblems. Besides he challenge frm he size f daa se, he uncerainy in he daa is als an issue, especially fr he image daa, where he uncerainy can cme frm he differences beween individual examples and disrins, e.g., pse, illuminain and nise. Direcly learning wih he riginal daa will lead a pr generalizain perfrmance since he meric ends verfi he uncerainy in he daa. By furher invesigaing he image daa, we find ha ms f riginal images can be bserved frm a much smaller se f clean laen examples wih differen disrins. The phenmenn is illusraed in Fig. 5. This bservain inspires us learn he meric wih laen examples in lieu f he riginal daa. The challenge is ha laen examples are unknwn and nly images wih uncerainies are available. In his wrk, we prpse a framewrk learn he disance meric and laen examples simulaneusly. I sufficienly explres he prperies f laen examples address he menined challenges. Firs, due he small size f laen examples, he sraegy f idenifying he acive se becmes affrdable when learning he meric. We adp i accelerae he learning prcedure via aviding he aemps n inacive cnsrains. Addiinally, cmpared wih he riginal daa, he uncerainy in laen examples decreases significanly. Cnsequenly, he meric direcly learned frm laen examples can fcus n he naure f he daa raher han he uncerainy in he daa. T furher imprve he rbusness, we adp he large margin prpery ha laen examples frm differen classes shuld be pushed away wih a daa dependen margin. Fig. 1 illusraes ha an apprpriae margin fr laen examples can als preserve he large margin fr he riginal daa. We cnduc he empirical sudy n benchmark image daa ses, including he challenging ImageNe daa se, demnsrae he efficacy and efficiency f he prpsed mehd. The res f he paper is rganized as fllws: Secin 2 summarizes he relaed wrk f DML. Secin 3 describes he deails f he prpsed mehd and Secin 4 summarizes he hereical analysis. Secin 5 cmpares he prpsed mehd he cnveninal DML mehds n he benchmark image daa ses. Finally, Secin 6 cncludes his wrk wih fuure direcins. 2. Relaed Wrk Many DML mehds have been prpsed in he pas decades [3, 17, 18] and cmprehensive surveys can be fund in [7, 19]. The represenaive mehds include Xing s mehd [18], ITML [3] and LMNN [17]. ITML learns a meric accrding pairwise cnsrains, where he disances beween pairs frm he same class shuld be smaller han a predefined hreshld and he disances beween pairs frm differen classes shuld be larger han anher predefined hreshld. LMNN is develped wih riple cnsrains and a meric is learned make sure ha pairs frm he same class are separaed frm he examples f differen classes wih a large margin. Cmpared wih pairwise cnsrains, riple cnsrains are mre flexible depic he lcal gemery. T handle he large number f cnsrains, sme mehds adp SGD r nline learning sample ne cnsrain r a mini-bach f cnsrains a each ierain [2, 10]. OA- SIS [2] randmly samples ne riple cnsrain a each ierain and cmpues he unbiased gradien accrdingly. When he size f he acive se is small, hese mehds require exremely large number f ierains imprve he mdel. Oher mehds ry explre he cncep f he acive se. LMNN [17] prpses learn he meric effecively a each ierain by cllecing an acive se ha cnsiss f cnsrains vilaed by he curren meric wihin he k-neares neighbrs fr each example. Hwever, i requires O(n 2 ) bain he apprpriae acive se. Besides he research abu cnveninal DML, deep meric learning has araced much aenin recenly [9, 13, 15, 16]. These sudies als indicae ha sampling acive riples is essenial fr acceleraing he cnvergence. FaceNe [15] keeps a large size f mini-bach and searches hard cnsrains wihin a mini-bach. LefedSruc [16] generaes he mini-bach wih he randmly seleced psiive examples and he crrespnding hard negaive examples. Prxy-NCA [9] adps prxy examples reduce he size f riple cnsrains. Once an anchr example is given, he similar and dissimilar examples will be searched wihin he se f prxies. In his wrk we prpse learn he meric nly wih laen examples which can dramaically reduce he cmpuainal cs f baining he acive se. Besides, he riangle inequaliy dse n hld fr he squared disance, which makes ur analysis significanly differen frm he exising wrk. 3. Margin Preserving Meric Learning Given a raining se {(x i, y i ) i = 1,, n}, where x i R d is an example and y i is he crrespnding label, DML

3 aims learn a gd disance meric such ha x i, x j, x k D 2 M (x i, x k ) D 2 M (x i, x j ) 1 where x i and x j are frm he same class and x k is differen. Given he disance meric M S+ d d, he squared disance is defined as D 2 M (x i, x j ) = (x i x j ) M(x i x j ) where S d d + denes he se f d d psiive semi-definie (PSD) marices. Fr he large-scale image daa se, we assume ha each bserved example is frm a laen example wih cerain zer mean disrins, i.e., i, E[x i ] = z :f(i)= where f( ) prjecs he riginal daa is crrespnding laen example. Then, we cnsider he expeced disance [20] beween bserved daa and he bjecive is learn a meric such ha x i, x j, x k E[D 2 M (x i, x k )] E[D 2 M (x i, x j )] 1 (1) Le z, z p and z q dene laen examples f x i, x j and x k respecively. Fr he disance beween examples frm he same class, we have E[D 2 M (x i, x j )] = E[(x i z + z ) M(x i z + z )] + E[(x j z p + z p ) M(x j z p + z p )] E[2x i Mx j ] = D 2 M (z, z p ) + E[D 2 M (x i, z )] + E[D 2 M (x j, z p )] = D 2 M (z, z p ) + 2E[D 2 M (x i, z )] (2) The las equain is due he fac ha x i and x j are i.i.d, since hey are frm he same class. By applying he same analysis fr he dissimilar pair, we have E[D 2 M (x i, x k )] = D 2 M (z, z + E[D 2 M (x i, z )] + E[D 2 M (x k, z ] D 2 M (z, z + E[D 2 M (x i, z )] (3) The inequaliy is because ha M is a PSD marix. Cmbining Eqns. 2 and 3, we find ha he difference beween he disances in he riginal riple can be lwer bunded by hse in he riple cnsising f laen examples E[D 2 M (x i, x k )] E[D 2 M (x i, x j )] D 2 M (z, z D 2 M (z, z p ) E[D 2 M (x i, z )] Therefre, he meric can be learned wih he cnsrains defined n laen examples such ha z, z p, z q D 2 M (z, z D 2 M (z, z p ) 1+E[D 2 M (x i, z )] Once he meric is bserved, he margin fr he expeced disances beween riginal daa (i.e., as in Eqn. 1) is als guaraneed. Cmpared wih he riginal cnsrains, he margin beween laen examples is increased by he facr f E[DM 2 (x i, z )]. This erm indicaes he expeced disance beween he riginal daa and is crrespnding laen example. I means ha he igher a lcal cluser is, he less a margin shuld be increased. Furhermre, each class akes a differen margin, which depends n he disribuin f he riginal daa and makes i mre flexible han a glbal margin. Wih he se f riples {z, z p, z q}, he pimizain prblem can be wrien as min L(M, z) = l(z, z p, z q; M) M S d d +, M F δ,z R d m where m n is he number f laen examples. We add a cnsrain fr he Frbenius nrm f he learned meric preven i frm verfiing. l( ) is he lss funcin and he hinge lss is applied in his wrk. l(z, z p, z q; M) = [1 + E[D 2 M (x i, z )] (D 2 M (z, z q) D 2 M (z, z p))] + This prblem is hard slve since bh he meric and laen examples are he variables be pimized. Therefre, we prpse slve i in an alernaing way and he deailed seps are demnsraed belw Updae z wih Upper Bund When fixing M k 1, he subprblem a he k-h ierain becmes min L(M k 1, z) = [ 1 + E[DM 2 z k 1 (x i, z )] }{{} a ] (DM 2 k 1 (z, z q) DM 2 k 1 (z, z p)) (4) }{{} + b The variable z appears in bh he erm f margin a and he erm f he riple difference b, which makes i hard pimize direcly. Our sraegy is find an apprpriae upper bund fr he riginal prblem and slve he simple prblem insead. Therem 1. The funcin L(M k 1, z) can be upper bunded by he series f funcins r F r(z). Fr he r- h class, we have F r (z) = c 1 E[DM 2 k 1 (x i, z )]+c 2 +c 3 DM 2 k 1 (z, z k 1 ) where c 1, c 2 and c 3 are cnsans and r F r(z k 1 ) = L(M k 1, z k 1 ).

4 The deailed prf can be fund in Secin 4. Afer remving he cnsan erms and rearrange he cefficiens, pimizing F r (z) is equivalen pimizing he fllwing prblem min Fr (z) = (5) z R d mr,µ:µi, {0,1}, µi,=1 µ i, DM 2 k 1 (x i, z ) + γ DM 2 k 1 (z, z k 1 ) i:y(i)=r where µ denes he membership ha assigns a laen example fr each riginal example. Till nw, i shws ha he riginal bjecive L(M k 1, z) can be upper bunded by r F r(z). Minimizing he upper bund is similar k-means bu wih he disance defined n he meric M k 1. S we can slve i by he sandard EM algrihm. When fixing µ, laen examples can be updaed by he clsed-frm sluin 1, z = i µ i, + γ ( i µ i, x i + γz k 1 ) (6) When fixing z, µ jus assigns each riginal example is neares laen example wih he disance defined n he meric M k 1 { 1 = arg min DM 2 i, µ i, = (x k 1 i, z ) (7) 0.w. Alg. 1 summarizes he mehd fr slving F r (z). Algrihm 1 Algrihm f Updaing z Inpu: daa se {X, Y }, z k 1, M k 1, γ and S Iniialize z = z k 1 fr s = 1 S d Fix z and bain he assignmen µ as in Eqn. 7 Fix µ and updae z as in Eqn. 6 end fr reurn z k = z 3.2. Updae M wih Upper Bund When fixing z k a he k-h ierain, he subprblem becmes min M S d d + L(M, z k ) = (8) [1 + E[DM 2 (x i, z )] (DM 2 (z }{{}, z q) DM 2 (z, z p)) }{{} a b where M als appears in muliple erms. Wih he similar prcedure, an upper bund can be fund make he pimizain simpler. ] + Therem 2. The funcin L(M, z k ) can be upper bunded by he funcin H(M) which is H(M) = λ 2 M M k 1 2 F + [ 1 + E[DM 2 k 1 (x i, z )] ] (DM 2 (z, z q) DM 2 (z, z p)) + where λ is a cnsan and H(M k 1 ) = L(M k 1, z k ). Minimizing H(M) is a sandard DML prblem. Since he number f laen examples z k is small, many exising DML mehds can handle he prblem well. In his wrk we slve he prblem by SGD bu sample ne epch acive cnsrains a each sage. The acive cnsrains cnain he riples f z k ha incur he hinge lss wih he disance defined n M k 1. This sraegy enjys he efficiency f SGD and he efficacy f learning wih he acive se. T furher imprve he efficiency, ne prjecin paradigm is adped avid he expensive PSD prjecin which css O(d 3 ). I perfrms he PSD prjecin nce a he end f he learning algrihm and shws be effecive in many applicains [2, 11]. Finally, since he prblem is srngly cnvex, we apply he α-suffix averaging sraegy, which averages he sluins ver he las several ierains, bain he pimal cnvergence rae [12]. The cmplee apprach fr baining M k is shwn in Alg. 2. Algrihm 2 Algrihm f Updaing M Inpu: daa se {X, Y }, z k, M k 1, δ, λ and S Iniialize M 0 = M k 1 Sample ne epch acive cnsrains A accrding z k and M k 1 fr s = 1 S d Randmly sample ne cnsrain frm A Cmpue he schasic gradien g = H(M) Updae he meric as M s = M s 1 1 λs g Check he Frbenius nrm M s = Π δ (M s) end fr Prjec he learned marix n he PSD cne M k = Π P SD ( 2 S S s=s/2+1 M s) reurn M k Alg. 3 summarizes he prpsed margin preserving meric learning framewrk. Differen frm he sandard alernaing mehd, we nly pimize he upper bund fr each subprblem. Hwever, he mehd cnverges as shwn in he fllwing herem. Therem 3. Le (z k 1, M k 1 ) and (z k, M k ) dene he resuls bained by applying he algrihm in Alg. 3 a (k 1)-h and k-h ierains respecively. Then, we have L(z k, M k ) L(z k 1, M k 1 ) which means he prpsed mehd can cnverge.

5 Algrihm 3 Margin Preserving Meric Learning (MaPML) Inpu: daa se {X, Y }, δ, m, γ, λ and K Iniialize M 0 = I fr k = 1 K d Fix M k 1 and bain laen examples z k by Alg. 1 Fix z k and updae he meric M k by Alg. 2 end fr reurn M K and z K Cmpuainal Cmplexiy The prpsed mehd cnsiss f w pars: baining laen examples and meric learning. Fr he frmer ne, he cs is linear in he number f laen examples and riginal examples as O(mn). Fr he laer ne, he cs f sampling an acive se dminaes he learning prcedure. Since he number f ierains is fixed, he cmplexiy f sampling becmes min{o(sm), O(m 2 )}. Therefre, he whle algrihm can be linear in he number f laen examples. Ne ha he efficiency can be furher imprved wih disribued cmpuing since many cmpnens f MaPML can be implemened in parallel. Fr example, when updaing z, each class is independen and all subprblems can be slved simulaneusly. 4. Thereical Analysis 4.1. Prf f Therem 1 Prf. Firs, fr he disance f he dissimilar pair in erm b f Eqn. 4, we have DM 2 (z, z = DM 2 (z k 1, zq k 1 ) + DM 2 (z, z k 1 ) + 2(z z k 1 ) M(z k 1 + DM 2 (z q, z k 1 2(z q z k 1 M(z k 1 2(z z k 1 ) M(z q z k 1 DM 2 (z k 1, z k 1 2D M (z, z k 1 2D M (z q, z k 1 D M (z k 1, z k 1 z k 1 z k 1 )D M (z k 1, z k 1 where z k 1 are laen examples frm he las ierain. We le M dene M k 1 in his prf fr simpliciy. The inequaliy is frm ha M is a PSD marix and can be decmpsed as M = LL. Then i is bained by applying he Cauchy-Schwarz inequaliy. Wih he assumpins ha, D M (z, z k 1 ) is sufficienly large and D M (z k 1, z k 1 is bunded by a cnsan c 2, he inequaliy can be simplified as D 2 M (z, z (9) D 2 M (z k 1, z k 1 ) cdm 2 (z, z k 1 ) cdm 2 (z q, z k 1 q The assumpin is easy verify since D M (z k 1, z k 1 ) z k 1 q z k 1 q 2 2 M k 1 2 Ne ha M k 1 2 M k 1 F δ and z is in he cnvex hull f he riginal daa, and he cnsan c can be se as c = 8δ max i x i 2 2. Wih he similar prcedure, we have he bund fr he disance f he similar pair as DM 2 (z, z p ) DM 2 (z k 1, z k 1 p ) (10) +(c + 2)DM 2 (z, z k 1 ) + (c + 2)DM 2 (z p, z k 1 p ) Taking Eqns. 9 and 10 back he riginal funcin L(M k 1, z) and using he prpery f he hinge lss, he riginal ne can be upper bunded by G(z) = [1 + E[DM 2 (x i, z )] (DM 2 (z :k 1, z :k 1 DM 2 (z :k 1, z :k 1 p ))] + + c 3 m D 2 M (z, z k 1 ) where c 3 = O(T c) is a cnsan. By invesigaing he srucure f his prblem, we find ha each class is independen in he pimizain prblem and he subprblem fr he r-h class can be wrien as min z R d mr G r (z) = + c 3 :y(z )=r :y(z )=r [E[D 2 M (x i, z )] + c ] + D 2 M (z, z k 1 ) where m r is he number f laen examples fr he r-h class and c is a cnsan as c = 1 (D 2 M (z :k 1, z :k 1 DM 2 (z :k 1, z :k 1 p )) Nex we ry upper bund he hinge lss in G r (z) wih a linear funcin in he inerval f [c, E[DM 2 (x i, z k 1 )]+c ], where he hinge lss incurred by he pimal sluin z k is guaraneed be in i. Figure 2. Illusrain f bunding he hinge lss. The hinge lss beween [c, α + c ] is upper bunded by he linear funcin dened by he red line. Le α = E[DM 2 (x i, z k 1 )], which is he expeced disance beween he riginal daa f he r-h class and he crrespnding laen examples frm he las ierain, and β be a cnsan sufficienly large as β min c

6 Then, fr each acive hinge lss (i.e., α + c > 0), if we have E[D 2 M (x i, z )] α (11) [E[DM 2 (x i, z )] + c ] + α + c α + c + β (E[D2 M (x i, z )] + c + β) Fig. 2 illusraes he linear funcin ha can bund he hinge lss and he prf is sraighfrward. We will shw ha he cndiin in Eqn. 11 can be saisfied hrughu he algrihm laer. Wih he upper bund f he hinge lss, G r (z) can be bunded by F r (z) = c 1 E[D 2 M (x i, z )] + c 2 + c 3 where and c 2 = c 1 = :y(z )=r :y(z )=r α + c α + c + β I(α + c ) D 2 M (z, z k 1 ) α + c α + c + β (c + β)i(α + c ) { 1 ν > 0 I( ) is an indicar funcin as I(ν) = 0.w. Finally, we check he cndiin in Eqn. 11. Le z k dene laen examples bained by pimizing F(z) wih Alg. 1. Since we use z k 1 as he saring pin pimize F r (z), i is bvius ha F r (z k ) F r (z k 1 ) A he same ime, we have DM 2 (z k, z k 1 ) DM 2 (z k 1, z k 1 ) = 0 I is bserved ha Eqn. 11 is saisfied by cmbining hese inequaliies Prf f Therem 2 Prf. Fr he erm a in Eqn. 8, we have E[D 2 M (x i, z )] = E[D 2 M k 1 (x i, z ) + (x i z ) (M M k 1 )(x i z )] E[DM 2 k 1 (x i, z )] + max x i z 2 2 M M k 1 F i E[DM 2 k 1 (x i, z )] + c M M k 1 2 F where we assume ha M M k 1 F is sufficienly large and c is a cnsan which has max i x i z 2 2 c and can be se as c = 4 max i x i 2 2. Therefre, he riginal funcin L(M, z k ) can be upper bunded by H(M) = λ 2 M M k 1 2 F + [ 1 + E[DM 2 k 1 (x i, z )] ] (DM 2 (z, z q) DM 2 (z, z p)) + where λ = O(T c) Prf f Therem 3 Prf. When fixing M k 1 a he k-h ierain, we have L(M k 1, z k ) r r When fixing z k, we have G r (z k ) r F r (z k 1 ) = L(M k 1, z k 1 ) F r (z k ) L(M k, z k ) H(M k ) H(M k 1 ) = L(M k 1, z k ) Therefre, afer each ierain, we have L(M k, z k ) L(M k 1, z k 1 ) Since he value f L( ) is bunded, he sequence will cnverge afer a finie number f ierains. 5. Experimens We cnduc he empirical sudy n fur benchmark image daa ses. 3-neares neighbr classifier is applied verify he efficacy f he learned merics frm differen mehds. The mehds in he cmparisn are summarized as fllws. Euclid: 3-NN wih Euclidean disance. LMNN [17]: he sae-f-he-ar DML mehd ha idenifies a se f acive riples wih he curren meric a each ierain. The acive riples are searched wihin 3-neares neighbrs fr each example. OASIS [2]: an nline DML mehd ha receives ne randm riple a each ierain. I nly updaes he meric when he riple cnsrain is acive. HR-SGD [10]: ne f he ms efficien DML mehds wih SGD. We adp he versin ha randmly samples a mini-bach f riples a each ierain in he cmparisn. Afer sampling, a Bernulli randm variable is generaed decide if updaing he curren meric r n. Wih he PSD prjecin, i guaranees ha he learned meric is in he PSD cne a each ierain.

7 MaPML τ : he prpsed mehd ha learns he meric and laen examples simulaneusly, where τ denes he rai beween he number f laen examples and he number f riginal nes τ% = m n Differen frm her mehds, 3-NN is implemened wih laen examples as reference pins. The mehd ha akes 3-NN wih riginal daa is referred as MaPML τ -O. The parameers f OASIS, HR-SGD and MaPML are searched in {10 i : i = 3,, 3}. The size f mini-bach in HR-SGD is se be 10 as suggesed [10]. T rain he mdel sufficienly, he number f ierains fr LMNN is se be 10 3 while he number f randmly sampled riples is 10 5 fr OASIS and HR-SGD. The number f ierains fr MaPML is se as K = 10 while he number f maximal ierains fr slving M k in he subprblem is se as S = 10 4, which rughly has he same number f riples as OASIS and HR-SGD. All experimens are implemened n a server wih 96 GB memry and 2 Inel Xen E CPUs. Average resuls wih sandard deviain ver 5 rails are repred MNIST Firs, we evaluae he perfrmance f differen algrihms n MNIST [8]. I cnsiss f 60, 000 handwrien digi images fr raining and 10, 000 images fr es. There are 10 classes in he daa se, which are crrespnding he digis 0-9. Each example is a grayscale image which leads he 784-dimensinal feaures and hey are nrmalized he range f [0, 1]. Tes errr (%) Euclid LMNN OASIS HR-SGD MaPML MaPML-O 5% 10% 15% 20% 25% Rai f #laen examples = Tes errr (%) Sandard deviain f Gaussian nise (a) Cmparisn f errr rae (b) Cmparisn wih Gaussian nise Figure 3. Cmparisns n MNIST Euclid LMNN MaPML10-O MaPML10 Fig. 3 (a) cmpares he perfrmance f differen merics n he es se. Fr MaPML, we vary he rai f laen examples frm 5% 25%. Firs f all, I is bvius ha he merics learned wih he acive se uperfrm hse frm randm riples. I cnfirms ha he sraegy f sampling riples randmly can n explre he daa se sufficienly due he exremely large number f riples. Secndly, he perfrmance f MaPML 10 -O is cmparable wih LMNN, which shws ha he prpsed mehd can learn a gd meric wih nly a small amun f laen examples (i.e., 10%). Finally, bh MaPML and MaPML-O wrk well wih he meric bained by MaPML, which verifies ha he learned meric can preserve he large margin prpery fr bh he riginal and laen daa. Ne ha when he number f laen examples is small, he perfrmance f k- NN wih laen examples is slighly wrse han ha wih he whle raining se. Hwever, k-nn wih laen examples can be mre rbus in real-wrld applicains. T demnsrae he rbusness, we cnduc anher experimen ha randmly inrduces he zer mean Gaussian nise (i.e., N (0, σ 2 )) each pixel f he riginal raining images. The sandard deviain f he Gaussian nise is varied in he range f [50/255, 250/255] and τ is fixed as 10. Fig. 3 (b) summarizes he resuls. I shws ha MaPML 10 has he cmparable perfrmance as MaPML 10 - O and LMNN when he nise level is lw. Hwever, wih he increasing f he nise, he perfrmance f LMNN drps dramaically. This can be inerpreed by he fac ha he meric learned wih he riginal daa has been misled by he nisy infrmain. In cnras, he errrs made by MaPML and MaPML-O increase mildly and i demnsraes ha he learned meric is mre rbus han he ne learned frm he riginal daa. MaPML perfrms bes amng all mehds and i is due he reasn ha he uncerainy in laen examples are much less han ha in he riginal nes. I implies ha k-nn wih laen examples is mre apprpriae fr real-wrld applicains wih large uncerainy. CPU Time (secnds) OASIS HR-SGD MaPML10 LMNN MNIST CIFAR-10 CIFAR #Ierains: k (a) CPU ime fr raining (b) Cnvergence curve f MaPML Figure 4. Illusrain f he efficiency f he prpsed mehd. Then, we cmpare he CPU ime cs by differen algrihms evaluae he efficiency. The resuls can be fund in Fig. 4 (a). Firs, as expeced, all algrihms wih SGD are mre efficien han LMNN, which has cmpue he full gradien frm he redefined acive se a each ierain. Mrever, he running ime f MaPML 10 is cmparable ha f HR-SGD, which shws he efficiency f MaPML wih he small se f laen examples. Ne ha OASIS has he exremely lw cs, since i allws he inernal meric be u f he PSD cne. Fig. 4 (b) illusraes he cnvergence curve f MaPML and shws ha he prpsed mehd cnverges fas in pracice. Finally, since we apply he prpsed mehd he rig- kzk! zk!1k 2 F

8 inal pixel feaures direcly, he learned laen examples can be recvered as images. Fig. 5 illusraes he learned laen examples and he crrespnding examples in he riginal raining se. I is bvius ha he riginal examples are frm laen examples wih differen disrins as claimed. Figure 5. Illusrain f he learned laen examples and crrespnding riginal examples frm MNIST. The lef clumn indicaes laen examples while five riginal images frm each crrespnding cluser are n he righ CIFAR-10 & CIFAR-100 CIFAR-10 cnains 10 classes wih 50, 000 clr images f size fr raining and 10, 000 images fr es. CIFAR-100 has he same number f images in raining and es bu fr 100 classes [6]. Since deep learning algrihms shw he verwhelming perfrmance n hese daa ses, we adp ResNe18 [4] in Caffe [5], which is pre-rained n ImageNe ILSVRC 2012 daa se [14], as he feaure exracr and each image is represened by a 512-dimensinal feaure vecr. Table 1. Cmparisn f errr rae (%) n CIFAR-10 and CIFAR Mehds CIFAR-10 CIFAR-100 Euclid OASIS ± ± 0.21 HR-SGD ± ± 0.19 LMNN ± ± 0.13 MaPML 10 -O ± ± 0.15 MaPML ± ± 0.16 Table 1 summarizes errr raes f mehds in he cmparisn. Firs, we have he same bservain as n MNIST, where he perfrmance f mehds adping acive riples is much beer han ha f he mehds wih randmly sampled riples. Differen frm MNIST, MaPML 10 uperfrms LMNN n bh f he daa ses. I is because ha images in hese daa ses describe naural bjecs which cnain much mre uncerainy han digis in MNIST. Finally, he perfrmance f MaPML 10 -O is superir ver OASIS and HR-SGD, which shws he learned meric can wrk well wih he riginal daa represened by deep feaures. I cnfirms ha he large margin prpery is preserved even fr he riginal daa ImageNe Finally, we demnsrae ha he prpsed mehd can handle he large-scale daa se wih ImageNe. ImageNe ILSVRC 2012 cnsiss f 1, 281, 167 raining images and 50, 000 validain daa. The same feaure exracin prcedure as abve is applied fr each image. Given he large number f raining daa, we increase he number f riples fr OASIS and HR-SGD Crrespndingly, he number f maximal ierains fr slving he subprblem in MaPML is als raised Table 2. Cmparisn f errr rae (%) n ImageNe. Mehds Tes errr (%) Euclid OASIS ± 0.08 HR-SGD ± 0.08 MaPML 5 -O ± 0.03 MaPML ± 0.09 LMNN des n finish he raining afer 24 hurs s he resul is n repred fr i. In cnras, MaPML bains he meric wihin abu ne hur. The perfrmance f available mehds can be fund in Table 2. Since ResNe18 is rained n ImageNe, he exraced feaures are pimized fr his daa se and i is hard furher imprve he perfrmance. Hwever, wih laen examples, MaPML can furher reduce he errr rae by 1.7%. I indicaes ha laen examples wih lw uncerainy are mre apprpriae fr he large-scale daa se as he reference pins. Ne ha he small number f reference pins will als accelerae he es phase. Fr example, i css 0.15s predic he label f an image wih he riginal se while he cs is nly 0.007s if evaluaing wih laen examples. I makes MaPML wih laen examples a penial mehd fr real-ime applicains. 6. Cnclusin In his wrk, we prpse a framewrk learn he disance meric and laen examples simulaneusly. By learning frm a small se f clean laen examples, MaPML can sample he acive riples efficienly and he learning prcedure is rbus he uncerainy in he real-wrld daa. Mrever, MaPML can preserve he large margin prpery fr he riginal daa when learning merely wih laen examples. The empirical sudy cnfirms he efficacy and efficiency f MaPML. In he fuure, we plan evaluae MaPML n differen asks (e.g., infrmain rerieval) and differen ypes f daa. Besides, incrpraing he prpsed sraegy deep meric learning is als an aracive direcin. I can accelerae he learning fr deep embedding and he resuling laen examples may furher imprve he perfrmance.

9 References [1] S. Byd and L. Vandenberghe. Cnvex Opimizain. Cambridge Universiy Press, New Yrk, NY, USA, [2] G. Chechik, V. Sharma, U. Shali, and S. Bengi. Large scale nline learning f image similariy hrugh ranking. JMLR, 11: , , 2, 4, 6 [3] J. V. Davis, B. Kulis, P. Jain, S. Sra, and I. S. Dhilln. Infrmain-hereic meric learning. In ICML, pages , , 2 [4] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning fr image recgniin. In CVPR, pages , [5] Y. Jia, E. Shelhamer, J. Dnahue, S. Karayev, J. Lng, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Cnvluinal archiecure fr fas feaure embedding. In ACM MM, pages , [6] A. Krizhevsky. Learning muliple layers f feaures frm iny images [7] B. Kulis. Meric learning: A survey. Fundains and Trends in Machine Learning, 5(4): , [8] Y. LeCun, L. Bu, Y. Bengi, and P. Haffner. Gradien-based learning applied dcumen recgniin. Prceedings f he IEEE, 86(11): , [9] Y. Mvshviz-Aias, A. Tshev, T. K. Leung, S. Iffe, and S. Singh. N fuss disance meric learning using prxies. In ICCV, pages , [10] Q. Qian, R. Jin, J. Yi, L. Zhang, and S. Zhu. Efficien disance meric learning by adapive sampling and mini-bach schasic gradien descen (SGD). ML, 99(3): , , 6, 7 [11] Q. Qian, R. Jin, S. Zhu, and Y. Lin. Fine-grained visual caegrizain via muli-sage meric learning. In CVPR, pages , [12] A. Rakhlin, O. Shamir, and K. Sridharan. Making gradien descen pimal fr srngly cnvex schasic pimizain. In ICML, [13] O. Rippel, M. Paluri, P. Dllár, and L. D. Burdev. Meric learning wih adapive densiy discriminain. In ICLR, [14] O. Russakvsky, J. Deng, H. Su, J. Krause, S. Saheesh, S. Ma, Z. Huang, A. Karpahy, A. Khsla, M. S. Bernsein, A. C. Berg, and F. Li. Imagene large scale visual recgniin challenge. IJCV, 115(3): , [15] F. Schrff, D. Kalenichenk, and J. Philbin. Facene: A unified embedding fr face recgniin and clusering. In CVPR, pages , [16] H. O. Sng, Y. Xiang, S. Jegelka, and S. Savarese. Deep meric learning via lifed srucured feaure embedding. In CVPR, pages , [17] K. Q. Weinberger and L. K. Saul. Disance meric learning fr large margin neares neighbr classificain. JMLR, 10: , , 2, 6 [18] E. P. Xing, A. Y. Ng, M. I. Jrdan, and S. J. Russell. Disance meric learning wih applicain clusering wih side-infrmain. In NIPS, pages , , 2 [19] L. Yang and R. Jin. Disance meric learning: a cmprehensive survery [20] H. Ye, D. Zhan, X. Si, and Y. Jiang. Learning mahalanbis disance meric: Cnsidering insance disurbance helps. In IJCAI, pages ,

MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.265/15.070J Fall 2013 Lecture 15 10/30/2013. Ito integral for simple processes

MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.265/15.070J Fall 2013 Lecture 15 10/30/2013. Ito integral for simple processes MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.65/15.7J Fall 13 Lecure 15 1/3/13 I inegral fr simple prcesses Cnen. 1. Simple prcesses. I ismery. Firs 3 seps in cnsrucing I inegral fr general prcesses 1 I inegral

More information

Productivity changes of units: A directional measure of cost Malmquist index

Productivity changes of units: A directional measure of cost Malmquist index Available nline a hp://jnrm.srbiau.ac.ir Vl.1, N.2, Summer 2015 Jurnal f New Researches in Mahemaics Science and Research Branch (IAU Prduciviy changes f unis: A direcinal measure f cs Malmquis index G.

More information

10.7 Temperature-dependent Viscoelastic Materials

10.7 Temperature-dependent Viscoelastic Materials Secin.7.7 Temperaure-dependen Viscelasic Maerials Many maerials, fr example plymeric maerials, have a respnse which is srngly emperaure-dependen. Temperaure effecs can be incrpraed in he hery discussed

More information

AP Physics 1 MC Practice Kinematics 1D

AP Physics 1 MC Practice Kinematics 1D AP Physics 1 MC Pracice Kinemaics 1D Quesins 1 3 relae w bjecs ha sar a x = 0 a = 0 and mve in ne dimensin independenly f ne anher. Graphs, f he velciy f each bjec versus ime are shwn belw Objec A Objec

More information

An application of nonlinear optimization method to. sensitivity analysis of numerical model *

An application of nonlinear optimization method to. sensitivity analysis of numerical model * An applicain f nnlinear pimizain mehd sensiiviy analysis f numerical mdel XU Hui 1, MU Mu 1 and LUO Dehai 2 (1. LASG, Insiue f Amspheric Physics, Chinese Academy f Sciences, Beijing 129, China; 2. Deparmen

More information

Convex Stochastic Duality and the Biting Lemma

Convex Stochastic Duality and the Biting Lemma Jurnal f Cnvex Analysis Vlume 9 (2002), N. 1, 237 244 Cnvex Schasic Dualiy and he Biing Lemma Igr V. Evsigneev Schl f Ecnmic Sudies, Universiy f Mancheser, Oxfrd Rad, Mancheser, M13 9PL, UK igr.evsigneev@man.ac.uk

More information

Brace-Gatarek-Musiela model

Brace-Gatarek-Musiela model Chaper 34 Brace-Gaarek-Musiela mdel 34. Review f HJM under risk-neural IP where f ( T Frward rae a ime fr brrwing a ime T df ( T ( T ( T d + ( T dw ( ( T The ineres rae is r( f (. The bnd prices saisfy

More information

The Components of Vector B. The Components of Vector B. Vector Components. Component Method of Vector Addition. Vector Components

The Components of Vector B. The Components of Vector B. Vector Components. Component Method of Vector Addition. Vector Components Upcming eens in PY05 Due ASAP: PY05 prees n WebCT. Submiing i ges yu pin ward yur 5-pin Lecure grade. Please ake i seriusly, bu wha cuns is wheher r n yu submi i, n wheher yu ge hings righ r wrng. Due

More information

5.1 Angles and Their Measure

5.1 Angles and Their Measure 5. Angles and Their Measure Secin 5. Nes Page This secin will cver hw angles are drawn and als arc lengh and rains. We will use (hea) represen an angle s measuremen. In he figure belw i describes hw yu

More information

Kinematics Review Outline

Kinematics Review Outline Kinemaics Review Ouline 1.1.0 Vecrs and Scalars 1.1 One Dimensinal Kinemaics Vecrs have magniude and direcin lacemen; velciy; accelerain sign indicaes direcin + is nrh; eas; up; he righ - is suh; wes;

More information

PRINCE SULTAN UNIVERSITY Department of Mathematical Sciences Final Examination Second Semester (072) STAT 271.

PRINCE SULTAN UNIVERSITY Department of Mathematical Sciences Final Examination Second Semester (072) STAT 271. PRINCE SULTAN UNIVERSITY Deparmen f Mahemaical Sciences Final Examinain Secnd Semeser 007 008 (07) STAT 7 Suden Name Suden Number Secin Number Teacher Name Aendance Number Time allwed is ½ hurs. Wrie dwn

More information

THE DETERMINATION OF CRITICAL FLOW FACTORS FOR NATURAL GAS MIXTURES. Part 3: The Calculation of C* for Natural Gas Mixtures

THE DETERMINATION OF CRITICAL FLOW FACTORS FOR NATURAL GAS MIXTURES. Part 3: The Calculation of C* for Natural Gas Mixtures A REPORT ON THE DETERMINATION OF CRITICAL FLOW FACTORS FOR NATURAL GAS MIXTURES Par 3: The Calculain f C* fr Naural Gas Mixures FOR NMSPU Deparmen f Trade and Indusry 151 Buckingham Palace Rad Lndn SW1W

More information

Coherent PSK. The functional model of passband data transmission system is. Signal transmission encoder. x Signal. decoder.

Coherent PSK. The functional model of passband data transmission system is. Signal transmission encoder. x Signal. decoder. Cheren PSK he funcinal mdel f passand daa ransmissin sysem is m i Signal ransmissin encder si s i Signal Mdular Channel Deecr ransmissin decder mˆ Carrier signal m i is a sequence f syml emied frm a message

More information

Lecture 4 ( ) Some points of vertical motion: Here we assumed t 0 =0 and the y axis to be vertical.

Lecture 4 ( ) Some points of vertical motion: Here we assumed t 0 =0 and the y axis to be vertical. Sme pins f erical min: Here we assumed and he y axis be erical. ( ) y g g y y y y g dwnwards 9.8 m/s g Lecure 4 Accelerain The aerage accelerain is defined by he change f elciy wih ime: a ; In analgy,

More information

An Introduction to Wavelet Analysis. with Applications to Vegetation Monitoring

An Introduction to Wavelet Analysis. with Applications to Vegetation Monitoring An Inrducin Wavele Analysis wih Applicains Vegeain Mniring Dn Percival Applied Physics Labrary, Universiy f Washingn Seale, Washingn, USA verheads fr alk available a hp://saff.washingn.edu/dbp/alks.hml

More information

Numerical solution of some types of fractional optimal control problems

Numerical solution of some types of fractional optimal control problems Numerical Analysis and Scienific mpuing Preprin Seria Numerical sluin f sme ypes f fracinal pimal cnrl prblems N.H. Sweilam T.M. Al-Ajmi R.H.W. Hppe Preprin #23 Deparmen f Mahemaics Universiy f Husn Nvember

More information

Visco-elastic Layers

Visco-elastic Layers Visc-elasic Layers Visc-elasic Sluins Sluins are bained by elasic viscelasic crrespndence principle by applying laplace ransfrm remve he ime variable Tw mehds f characerising viscelasic maerials: Mechanical

More information

RAPIDLY ADAPTIVE CFAR DETECTION BY MERGING INDIVIDUAL DECISIONS FROM TWO-STAGE ADAPTIVE DETECTORS

RAPIDLY ADAPTIVE CFAR DETECTION BY MERGING INDIVIDUAL DECISIONS FROM TWO-STAGE ADAPTIVE DETECTORS RAPIDLY ADAPIVE CFAR DEECION BY MERGING INDIVIDUAL DECISIONS FROM WO-SAGE ADAPIVE DEECORS Analii A. Knnv, Sung-yun Chi and Jin-a Kim Research Cener, SX Engine Yngin-si, 694 Krea kaa@ieee.rg; dkrein@nesx.cm;

More information

Unit-I (Feedback amplifiers) Features of feedback amplifiers. Presentation by: S.Karthie, Lecturer/ECE SSN College of Engineering

Unit-I (Feedback amplifiers) Features of feedback amplifiers. Presentation by: S.Karthie, Lecturer/ECE SSN College of Engineering Uni-I Feedback ampliiers Feaures eedback ampliiers Presenain by: S.Karhie, Lecurer/ECE SSN Cllege Engineering OBJECTIVES T make he sudens undersand he eec negaive eedback n he llwing ampliier characerisics:

More information

CHAPTER 7 CHRONOPOTENTIOMETRY. In this technique the current flowing in the cell is instantaneously stepped from

CHAPTER 7 CHRONOPOTENTIOMETRY. In this technique the current flowing in the cell is instantaneously stepped from CHAPTE 7 CHONOPOTENTIOMETY In his echnique he curren flwing in he cell is insananeusly sepped frm zer sme finie value. The sluin is n sirred and a large ecess f suppring elecrlye is presen in he sluin;

More information

and Sun (14) and Due and Singlen (19) apply he maximum likelihd mehd while Singh (15), and Lngsa and Schwarz (12) respecively emply he hreesage leas s

and Sun (14) and Due and Singlen (19) apply he maximum likelihd mehd while Singh (15), and Lngsa and Schwarz (12) respecively emply he hreesage leas s A MONTE CARLO FILTERING APPROACH FOR ESTIMATING THE TERM STRUCTURE OF INTEREST RATES Akihik Takahashi 1 and Seish Sa 2 1 The Universiy f Tky, 3-8-1 Kmaba, Megur-ku, Tky 153-8914 Japan 2 The Insiue f Saisical

More information

Impact Switch Study Modeling & Implications

Impact Switch Study Modeling & Implications L-3 Fuzing & Ordnance Sysems Impac Swich Sudy Mdeling & Implicains Dr. Dave Frankman May 13, 010 NDIA 54 h Annual Fuze Cnference This presenain cnsiss f L-3 Crprain general capabiliies infrmain ha des

More information

GMM Estimation of the Number of Latent Factors

GMM Estimation of the Number of Latent Factors GMM Esimain f he Number f aen Facrs Seung C. Ahn a, Marcs F. Perez b March 18, 2007 Absrac We prpse a generalized mehd f mmen (GMM) esimar f he number f laen facrs in linear facr mdels. he mehd is apprpriae

More information

Successive ApproxiInations and Osgood's Theorenl

Successive ApproxiInations and Osgood's Theorenl Revisa de la Unin Maemaica Argenina Vlumen 40, Niimers 3 y 4,1997. 73 Successive ApprxiInains and Osgd's Therenl Calix P. Caldern Virginia N. Vera de Seri.July 29, 1996 Absrac The Picard's mehd fr slving

More information

Motion Along a Straight Line

Motion Along a Straight Line PH 1-3A Fall 010 Min Alng a Sraigh Line Lecure Chaper (Halliday/Resnick/Walker, Fundamenals f Physics 8 h ediin) Min alng a sraigh line Sudies he min f bdies Deals wih frce as he cause f changes in min

More information

GAMS Handout 2. Utah State University. Ethan Yang

GAMS Handout 2. Utah State University. Ethan Yang Uah ae Universiy DigialCmmns@UU All ECAIC Maerials ECAIC Repsiry 2017 GAM Handu 2 Ehan Yang yey217@lehigh.edu Fllw his and addiinal wrs a: hps://digialcmmns.usu.edu/ecsaic_all Par f he Civil Engineering

More information

Lecture 3: Resistive forces, and Energy

Lecture 3: Resistive forces, and Energy Lecure 3: Resisive frces, and Energy Las ie we fund he velciy f a prjecile ving wih air resisance: g g vx ( ) = vx, e vy ( ) = + v + e One re inegrain gives us he psiin as a funcin f ie: dx dy g g = vx,

More information

Physics 111. Exam #1. September 28, 2018

Physics 111. Exam #1. September 28, 2018 Physics xam # Sepember 8, 08 ame Please read and fllw hese insrucins carefully: Read all prblems carefully befre aemping slve hem. Yur wrk mus be legible, and he rganizain clear. Yu mus shw all wrk, including

More information

Answers: ( HKMO Heat Events) Created by: Mr. Francis Hung Last updated: 21 September 2018

Answers: ( HKMO Heat Events) Created by: Mr. Francis Hung Last updated: 21 September 2018 nswers: (009-0 HKMO Hea Evens) reaed by: Mr. Francis Hung Las updaed: Sepember 08 09-0 Individual 6 7 7 0 Spare 8 9 0 08 09-0 8 0 0.8 Spare Grup 6 0000 7 09 8 00 9 0 0 Individual Evens I In hw many pssible

More information

Physics Courseware Physics I Constant Acceleration

Physics Courseware Physics I Constant Acceleration Physics Curseware Physics I Cnsan Accelerain Equains fr cnsan accelerain in dimensin x + a + a + ax + x Prblem.- In he 00-m race an ahlee acceleraes unifrmly frm res his p speed f 0m/s in he firs x5m as

More information

SMKA NAIM LILBANAT KOTA BHARU KELANTAN. SEKOLAH BERPRESTASI TINGGI. Kertas soalan ini mengandungi 7 halaman bercetak.

SMKA NAIM LILBANAT KOTA BHARU KELANTAN. SEKOLAH BERPRESTASI TINGGI. Kertas soalan ini mengandungi 7 halaman bercetak. Name : Frm :. SMKA NAIM LILBANAT 55 KOTA BHARU KELANTAN. SEKOLAH BERPRESTASI TINGGI PEPERIKSAAN PERCUBAAN SPM / ADDITIONAL MATHEMATICS Keras ½ Jam ½ Jam Unuk Kegunaan Pemeriksa Arahan:. This quesin paper

More information

The 37th International Physics Olympiad Singapore. Experimental Competition. Wednesday, 12 July, Sample Solution

The 37th International Physics Olympiad Singapore. Experimental Competition. Wednesday, 12 July, Sample Solution The 37h Inernainal Physics Olypiad Singapre Experienal Cpeiin Wednesday, July, 006 Saple Sluin Par a A skech f he experienal seup (n required) Receiver Raing able Gnieer Fixed ar Bea splier Gnieer Mvable

More information

independenly fllwing square-r prcesses, he inuiive inerpreain f he sae variables is n clear, and smeimes i seems dicul nd admissible parameers fr whic

independenly fllwing square-r prcesses, he inuiive inerpreain f he sae variables is n clear, and smeimes i seems dicul nd admissible parameers fr whic A MONTE CARLO FILTERING APPROACH FOR ESTIMATING THE TERM STRUCTURE OF INTEREST RATES Akihik Takahashi 1 and Seish Sa 2 1 The Universiy f Tky, 3-8-1 Kmaba, Megur-ku, Tky 153-8914 Japan 2 The Insiue f Saisical

More information

ON THE COMPONENT DISTRIBUTION COEFFICIENTS AND SOME REGULARITIES OF THE CRYSTALLIZATION OF SOLID SOLUTION ALLOYS IN MULTICOMPONENT SYSTEMS*

ON THE COMPONENT DISTRIBUTION COEFFICIENTS AND SOME REGULARITIES OF THE CRYSTALLIZATION OF SOLID SOLUTION ALLOYS IN MULTICOMPONENT SYSTEMS* METL 006.-5.5.006, Hradec nad Mravicí ON THE OMPONENT DISTRIUTION OEFFIIENTS ND SOME REGULRITIES OF THE RYSTLLIZTION OF SOLID SOLUTION LLOYS IN MULTIOMPONENT SYSTEMS* Eugenij V.Sidrv a, M.V.Pikunv b, Jarmír.Drápala

More information

Microwave Engineering

Microwave Engineering Micrwave Engineering Cheng-Hsing Hsu Deparmen f Elecrical Engineering Nainal Unied Universiy Ouline. Transmissin ine Thery. Transmissin ines and Waveguides eneral Sluins fr TEM, TE, and TM waves ; Parallel

More information

Index-based Most Similar Trajectory Search

Index-based Most Similar Trajectory Search Index-based Ms Similar rajecry Search Elias Frenzs, Ksas Grasias, Yannis hedridis Labrary f Infrmain Sysems eparmen f Infrmaics Universiy f Piraeus Hellas echnical Repr Series UNIPI-ISL-R-6- Nvember 6

More information

21.9 Magnetic Materials

21.9 Magnetic Materials 21.9 Magneic Maerials The inrinsic spin and rbial min f elecrns gives rise he magneic prperies f maerials è elecrn spin and rbis ac as iny curren lps. In ferrmagneic maerials grups f 10 16-10 19 neighbring

More information

Index-based Most Similar Trajectory Search

Index-based Most Similar Trajectory Search Index-based Ms Similar rajecry Search Elias Frenzs Ksas Grasias Yannis hedridis ep. f Infrmaics, Universiy f Piraeus, Greece ep. f Infrmaics, Universiy f Piraeus, Greece ep. f Infrmaics, Universiy f Piraeus,

More information

Nelson Primary School Written Calculation Policy

Nelson Primary School Written Calculation Policy Addiin Fundain Y1 Y2 Children will engage in a wide variey f sngs, rhymes, games and aciviies. They will begin relae addiin cmbining w grups f bjecs. They will find ne mre han a given number. Cninue develp

More information

Revelation of Soft-Switching Operation for Isolated DC to Single-phase AC Converter with Power Decoupling

Revelation of Soft-Switching Operation for Isolated DC to Single-phase AC Converter with Power Decoupling Revelain f Sf-Swiching Operain fr Islaed DC Single-phase AC Cnverer wih wer Decupling Nagisa Takaka, Jun-ichi Ih Dep. f Elecrical Engineering Nagaka Universiy f Technlgy Nagaka, Niigaa, Japan nakaka@sn.nagakau.ac.jp,

More information

Module 4. Analysis of Statically Indeterminate Structures by the Direct Stiffness Method. Version 2 CE IIT, Kharagpur

Module 4. Analysis of Statically Indeterminate Structures by the Direct Stiffness Method. Version 2 CE IIT, Kharagpur Mdle Analysis f Saically Indeerminae Srcres by he Direc Siffness Mehd Versin CE IIT, Kharagr Lessn The Direc Siffness Mehd: Temerare Changes and Fabricain Errrs in Trss Analysis Versin CE IIT, Kharagr

More information

Review of HAARP Experiment and Assessment of Ionospheric Effects

Review of HAARP Experiment and Assessment of Ionospheric Effects Third AL PI ympsium Kna, Hawaii Nvember 9-3, 009 Review f HAARP Experimen and Assessmen f Inspheric Effecs T. L. Ainswrh, Y. Wang, J.-. Lee, and K.-. Chen Naval Research Labrary, Washingn DC, UA CRR, Nainal

More information

Online Convex Optimization Example And Follow-The-Leader

Online Convex Optimization Example And Follow-The-Leader CSE599s, Spring 2014, Online Learning Lecure 2-04/03/2014 Online Convex Opimizaion Example And Follow-The-Leader Lecurer: Brendan McMahan Scribe: Sephen Joe Jonany 1 Review of Online Convex Opimizaion

More information

Internal Audit Report NGO: Tagore Society for Rural Development

Internal Audit Report NGO: Tagore Society for Rural Development 1 I N T E G R A T E D C O A S T A L Z O N E M A N A G E M E N T P R O J E C T W E S T B E N G A L N G O : T a g r e S c i e y f r R u r a l D e v e l p m e n C n e n s Secin I: Audi Scpe and Apprach:...

More information

Online Influence Maximization under Independent Cascade Model with Semi-Bandit Feedback

Online Influence Maximization under Independent Cascade Model with Semi-Bandit Feedback Online Influence Maximizain under Independen Cascade Mdel wih Semi-Bandi Feedback Zheng Wen Adbe Research zwen@adbe.cm Michal Valk SequeL eam, INRIA Lille - Nrd Eurpe michal.valk@inria.fr Branislav Kven

More information

Efficient and Fast Simulation of RF Circuits and Systems via Spectral Method

Efficient and Fast Simulation of RF Circuits and Systems via Spectral Method Efficien and Fas Simulain f RF Circuis and Sysems via Specral Mehd 1. Prjec Summary The prpsed research will resul in a new specral algrihm, preliminary simular based n he new algrihm will be subsanially

More information

Section 12 Time Series Regression with Non- Stationary Variables

Section 12 Time Series Regression with Non- Stationary Variables Secin Time Series Regressin wih Nn- Sainary Variables The TSMR assumpins include, criically, he assumpin ha he variables in a regressin are sainary. Bu many (ms?) ime-series variables are nnsainary. We

More information

Vehicle Arrival Models : Headway

Vehicle Arrival Models : Headway Chaper 12 Vehicle Arrival Models : Headway 12.1 Inroducion Modelling arrival of vehicle a secion of road is an imporan sep in raffic flow modelling. I has imporan applicaion in raffic flow simulaion where

More information

K The slowest step in a mechanism has this

K The slowest step in a mechanism has this CM 6 Generl Chemisry II Nme SLUTINS Exm, Spring 009 Dr. Seel. (0 pins) Selec he nswer frm he clumn n he righ h bes mches ech descripin frm he clumn n he lef. Ech nswer cn be used, ms, nly nce. E G This

More information

The lower limit of interval efficiency in Data Envelopment Analysis

The lower limit of interval efficiency in Data Envelopment Analysis Jurnal f aa nelpmen nalysis and ecisin Science 05 N. (05) 58-66 ailable nline a www.ispacs.cm/dea lume 05, Issue, ear 05 ricle I: dea-00095, 9 Pages di:0.5899/05/dea-00095 Research ricle aa nelpmen nalysis

More information

Thermal Analysis Validation for Different Design Tubes in a Heat Exchanger

Thermal Analysis Validation for Different Design Tubes in a Heat Exchanger Thermal Analysis Validain fr Differen Design Tubes in a Hea Exchanger Rshan. V. Marde, Ashk. J. Keche Deparmen f Mechanical Engineering, MIT, Aurangabad (M., India Aricle Inf Aricle hisry: Received 2 January

More information

A Generalized Schwarz Splitting Method Based on Hermite Collocation for Elliptic Boundary Value Problems

A Generalized Schwarz Splitting Method Based on Hermite Collocation for Elliptic Boundary Value Problems Purdue Universiy Purdue e-pubs Deparmen f Cmpuer Science Technical Reprs Deparmen f Cmpuer Science 1993 A Generalized Schwarz Spliing Mehd Based n Hermie Cllcain fr Ellipic Bundary Value Prblems Yu-Ling

More information

Finite time L 1 Approach for Missile Overload Requirement Analysis in Terminal Guidance

Finite time L 1 Approach for Missile Overload Requirement Analysis in Terminal Guidance Chinese Jurnal Aernauics 22(29) 413-418 Chinese Jurnal Aernauics www.elsevier.cm/lcae/cja Finie ime L 1 Apprach r Missile Overlad Requiremen Analysis in Terminal Guidance Ji Dengga*, He Fenghua, Ya Yu

More information

Matrix Versions of Some Refinements of the Arithmetic-Geometric Mean Inequality

Matrix Versions of Some Refinements of the Arithmetic-Geometric Mean Inequality Marix Versions of Some Refinemens of he Arihmeic-Geomeric Mean Inequaliy Bao Qi Feng and Andrew Tonge Absrac. We esablish marix versions of refinemens due o Alzer ], Carwrigh and Field 4], and Mercer 5]

More information

Chapter 3: Cluster Analysis

Chapter 3: Cluster Analysis Chapter 3: Cluster Analysis } 3.1 Basic Cncepts f Clustering 3.1.1 Cluster Analysis 3.1. Clustering Categries } 3. Partitining Methds 3..1 The principle 3.. K-Means Methd 3..3 K-Medids Methd 3..4 CLARA

More information

The Buck Resonant Converter

The Buck Resonant Converter EE646 Pwer Elecrnics Chaper 6 ecure Dr. Sam Abdel-Rahman The Buck Resnan Cnverer Replacg he swich by he resnan-ype swich, ba a quasi-resnan PWM buck cnverer can be shwn ha here are fur mdes f pera under

More information

Ramsey model. Rationale. Basic setup. A g A exogenous as in Solow. n L each period.

Ramsey model. Rationale. Basic setup. A g A exogenous as in Solow. n L each period. Ramsey mdel Rainale Prblem wih he Slw mdel: ad-hc assumpin f cnsan saving rae Will cnclusins f Slw mdel be alered if saving is endgenusly deermined by uiliy maximizain? Yes, bu we will learn a l abu cnsumpin/saving

More information

Optimization of Four-Button BPM Configuration for Small-Gap Beam Chambers

Optimization of Four-Button BPM Configuration for Small-Gap Beam Chambers Opimizain f Fur-Bun BPM Cnfigurain fr Small-Gap Beam Chamers S. H. Kim Advanced Phn Surce Argnne Nainal Larary 9700 Suh Cass Avenue Argnne, Illinis 60439 USA Asrac. The cnfigurain f fur-un eam psiin mnirs

More information

Stability of the SDDRE based Estimator for Stochastic Nonlinear System

Stability of the SDDRE based Estimator for Stochastic Nonlinear System 26 ISCEE Inernainal Cnference n he Science f Elecrical Engineering Sabiliy f he SDDRE based Esimar fr Schasic Nnlinear Sysem Ilan Rusnak Senir Research Fellw, RAFAEL (63, P.O.Bx 225, 322, Haifa, Israel.;

More information

Ensamble methods: Boosting

Ensamble methods: Boosting Lecure 21 Ensamble mehods: Boosing Milos Hauskrech milos@cs.pi.edu 5329 Senno Square Schedule Final exam: April 18: 1:00-2:15pm, in-class Term projecs April 23 & April 25: a 1:00-2:30pm in CS seminar room

More information

Money in OLG Models. 1. Introduction. Econ604. Spring Lutz Hendricks

Money in OLG Models. 1. Introduction. Econ604. Spring Lutz Hendricks Mne in OLG Mdels Ecn604. Spring 2005. Luz Hendricks. Inrducin One applicain f he mdels sudied in his curse ha will be pursued hrughu is mne. The purpse is w-fld: I prvides an inrducin he ke mdels f mne

More information

Machine Learning for Signal Processing Prediction and Estimation, Part II

Machine Learning for Signal Processing Prediction and Estimation, Part II Machine Learning fr Signal Prceing Predicin and Eimain, Par II Bhikha Raj Cla 24. 2 Nv 203 2 Nv 203-755/8797 Adminirivia HW cre u Sme uden wh g really pr mark given chance upgrade Make i all he way he

More information

23.2. Representing Periodic Functions by Fourier Series. Introduction. Prerequisites. Learning Outcomes

23.2. Representing Periodic Functions by Fourier Series. Introduction. Prerequisites. Learning Outcomes Represening Periodic Funcions by Fourier Series 3. Inroducion In his Secion we show how a periodic funcion can be expressed as a series of sines and cosines. We begin by obaining some sandard inegrals

More information

Speaker Adaptation Techniques For Continuous Speech Using Medium and Small Adaptation Data Sets. Constantinos Boulis

Speaker Adaptation Techniques For Continuous Speech Using Medium and Small Adaptation Data Sets. Constantinos Boulis Speaker Adapaion Techniques For Coninuous Speech Using Medium and Small Adapaion Daa Ses Consaninos Boulis Ouline of he Presenaion Inroducion o he speaker adapaion problem Maximum Likelihood Sochasic Transformaions

More information

STATE-SPACE MODELLING. A mass balance across the tank gives:

STATE-SPACE MODELLING. A mass balance across the tank gives: B. Lennox and N.F. Thornhill, 9, Sae Space Modelling, IChemE Process Managemen and Conrol Subjec Group Newsleer STE-SPACE MODELLING Inroducion: Over he pas decade or so here has been an ever increasing

More information

INFLUENCE OF WIND VELOCITY TO SUPPLY WATER TEMPERATURE IN HOUSE HEATING INSTALLATION AND HOT-WATER DISTRICT HEATING SYSTEM

INFLUENCE OF WIND VELOCITY TO SUPPLY WATER TEMPERATURE IN HOUSE HEATING INSTALLATION AND HOT-WATER DISTRICT HEATING SYSTEM Dr. Branislav Zivkvic, B. Eng. Faculy f Mechanical Engineering, Belgrade Universiy Predrag Zeknja, B. Eng. Belgrade Municipal DH Cmpany Angelina Kacar, B. Eng. Faculy f Agriculure, Belgrade Universiy INFLUENCE

More information

On the Resistance of an Infinite Square Network of Identical Resistors (Theoretical and Experimental Comparison)

On the Resistance of an Infinite Square Network of Identical Resistors (Theoretical and Experimental Comparison) On he esisance f an Infinie Square Newrk f Idenical esisrs (Thereical and Experimenal Cmparisn) J. H. Asad, A. Sakai,. S. Hiawi and J. M. Khalifeh Deparmen f Phsics, Universi f Jrdan, Amman-1194, Jrdan.

More information

Finish reading Chapter 2 of Spivak, rereading earlier sections as necessary. handout and fill in some missing details!

Finish reading Chapter 2 of Spivak, rereading earlier sections as necessary. handout and fill in some missing details! MAT 257, Handou 6: Ocober 7-2, 20. I. Assignmen. Finish reading Chaper 2 of Spiva, rereading earlier secions as necessary. handou and fill in some missing deails! II. Higher derivaives. Also, read his

More information

Ensamble methods: Bagging and Boosting

Ensamble methods: Bagging and Boosting Lecure 21 Ensamble mehods: Bagging and Boosing Milos Hauskrech milos@cs.pi.edu 5329 Senno Square Ensemble mehods Mixure of expers Muliple base models (classifiers, regressors), each covers a differen par

More information

A Dynamic Model of Economic Fluctuations

A Dynamic Model of Economic Fluctuations CHAPTER 15 A Dynamic Model of Economic Flucuaions Modified for ECON 2204 by Bob Murphy 2016 Worh Publishers, all righs reserved IN THIS CHAPTER, OU WILL LEARN: how o incorporae dynamics ino he AD-AS model

More information

Subject: Turbojet engines (continued); Design parameters; Effect of mass flow on thrust.

Subject: Turbojet engines (continued); Design parameters; Effect of mass flow on thrust. 16.50 Leure 19 Subje: Turbje engines (ninued; Design parameers; Effe f mass flw n hrus. In his haper we examine he quesin f hw hse he key parameers f he engine bain sme speified perfrmane a he design ndiins,

More information

Supplement for Stochastic Convex Optimization: Faster Local Growth Implies Faster Global Convergence

Supplement for Stochastic Convex Optimization: Faster Local Growth Implies Faster Global Convergence Supplemen for Sochasic Convex Opimizaion: Faser Local Growh Implies Faser Global Convergence Yi Xu Qihang Lin ianbao Yang Proof of heorem heorem Suppose Assumpion holds and F (w) obeys he LGC (6) Given

More information

Explaining Total Factor Productivity. Ulrich Kohli University of Geneva December 2015

Explaining Total Factor Productivity. Ulrich Kohli University of Geneva December 2015 Explaining Toal Facor Produciviy Ulrich Kohli Universiy of Geneva December 2015 Needed: A Theory of Toal Facor Produciviy Edward C. Presco (1998) 2 1. Inroducion Toal Facor Produciviy (TFP) has become

More information

Testing for a Single Factor Model in the Multivariate State Space Framework

Testing for a Single Factor Model in the Multivariate State Space Framework esing for a Single Facor Model in he Mulivariae Sae Space Framework Chen C.-Y. M. Chiba and M. Kobayashi Inernaional Graduae School of Social Sciences Yokohama Naional Universiy Japan Faculy of Economics

More information

RL Lecture 7: Eligibility Traces. R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 1

RL Lecture 7: Eligibility Traces. R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 1 RL Lecure 7: Eligibiliy Traces R. S. Suon and A. G. Baro: Reinforcemen Learning: An Inroducion 1 N-sep TD Predicion Idea: Look farher ino he fuure when you do TD backup (1, 2, 3,, n seps) R. S. Suon and

More information

GMM - Generalized Method of Moments

GMM - Generalized Method of Moments GMM - Generalized Mehod of Momens Conens GMM esimaion, shor inroducion 2 GMM inuiion: Maching momens 2 3 General overview of GMM esimaion. 3 3. Weighing marix...........................................

More information

Fractional Order Disturbance Observer based Robust Control

Fractional Order Disturbance Observer based Robust Control 201 Inernainal Cnference n Indusrial Insrumenain and Cnrl (ICIC) Cllege f Engineering Pune, India. May 28-30, 201 Fracinal Order Disurbance Observer based Rbus Cnrl Bhagyashri Tamhane 1, Amrua Mujumdar

More information

(1.1) (f'(z),y-x)>f(y)-f(x).

(1.1) (f'(z),y-x)>f(y)-f(x). PROCEEDINGS OF THE AMERICAN MATHEMATICAL SOCIETY Vlume 122, Number 4, December 1994 MEAN VALUE INEQUALITIES F. H. CLARKE AND YU. S. LEDYAEV (Cmmunicaed by Andrew Bruckner) Absrac. We prve a new ype f mean

More information

Random Walk with Anti-Correlated Steps

Random Walk with Anti-Correlated Steps Random Walk wih Ani-Correlaed Seps John Noga Dirk Wagner 2 Absrac We conjecure he expeced value of random walks wih ani-correlaed seps o be exacly. We suppor his conjecure wih 2 plausibiliy argumens and

More information

Expert Advice for Amateurs

Expert Advice for Amateurs Exper Advice for Amaeurs Ernes K. Lai Online Appendix - Exisence of Equilibria The analysis in his secion is performed under more general payoff funcions. Wihou aking an explici form, he payoffs of he

More information

Lecture II Simple One-Dimensional Vibrating Systems

Lecture II Simple One-Dimensional Vibrating Systems UIUC Physics 406 Acusical Physics f Music Lecure II Simple One-Dimensinal Vibraing Sysems One mehd f prducing a sund relies n a physical bjec (e.g. varius ypes f musical insrumens sringed and wind insrumens

More information

An recursive analytical technique to estimate time dependent physical parameters in the presence of noise processes

An recursive analytical technique to estimate time dependent physical parameters in the presence of noise processes WHAT IS A KALMAN FILTER An recursive analyical echnique o esimae ime dependen physical parameers in he presence of noise processes Example of a ime and frequency applicaion: Offse beween wo clocks PREDICTORS,

More information

Lyapunov Stability Stability of Equilibrium Points

Lyapunov Stability Stability of Equilibrium Points Lyapunv Stability Stability f Equilibrium Pints 1. Stability f Equilibrium Pints - Definitins In this sectin we cnsider n-th rder nnlinear time varying cntinuus time (C) systems f the frm x = f ( t, x),

More information

A Note on the Approximation of the Wave Integral. in a Slightly Viscous Ocean of Finite Depth. due to Initial Surface Disturbances

A Note on the Approximation of the Wave Integral. in a Slightly Viscous Ocean of Finite Depth. due to Initial Surface Disturbances Applied Mahemaical Sciences, Vl. 7, 3, n. 36, 777-783 HIKARI Ld, www.m-hikari.cm A Ne n he Apprximain f he Wave Inegral in a Slighly Viscus Ocean f Finie Deph due Iniial Surface Disurbances Arghya Bandypadhyay

More information

Lecture 2-1 Kinematics in One Dimension Displacement, Velocity and Acceleration Everything in the world is moving. Nothing stays still.

Lecture 2-1 Kinematics in One Dimension Displacement, Velocity and Acceleration Everything in the world is moving. Nothing stays still. Lecure - Kinemaics in One Dimension Displacemen, Velociy and Acceleraion Everyhing in he world is moving. Nohing says sill. Moion occurs a all scales of he universe, saring from he moion of elecrons in

More information

Method of Orthogonal Potentials Developed for the Analysis of TEM Mode Electromagnetic Resonators

Method of Orthogonal Potentials Developed for the Analysis of TEM Mode Electromagnetic Resonators 14-016-01-01_00 R.F. Ne #15 NSCL June 1, 005 Jhn incen Mehd f Orhgnal Penials Develed fr he Analysis f TEM Mde Elecrmagneic Resnars INTRODUCTION... DEELOPMENT... 3 E, H FIELD, ω... 4 SUMMARY EQUATIONS...

More information

Online Hashing. Long-Kai Huang, Qiang Yang, and Wei-Shi Zheng

Online Hashing. Long-Kai Huang, Qiang Yang, and Wei-Shi Zheng 1 Online Hashing Long-Kai Huang, Qiang Yang, and Wei-Shi Zheng Absrac Alhough hash funcion learning algorihms have achieved grea success in recen years, mos exising hash models are off-line, which are

More information

KINEMATICS IN ONE DIMENSION

KINEMATICS IN ONE DIMENSION KINEMATICS IN ONE DIMENSION PREVIEW Kinemaics is he sudy of how hings move how far (disance and displacemen), how fas (speed and velociy), and how fas ha how fas changes (acceleraion). We say ha an objec

More information

Comparing Means: t-tests for One Sample & Two Related Samples

Comparing Means: t-tests for One Sample & Two Related Samples Comparing Means: -Tess for One Sample & Two Relaed Samples Using he z-tes: Assumpions -Tess for One Sample & Two Relaed Samples The z-es (of a sample mean agains a populaion mean) is based on he assumpion

More information

ELEG 205 Fall Lecture #10. Mark Mirotznik, Ph.D. Professor The University of Delaware Tel: (302)

ELEG 205 Fall Lecture #10. Mark Mirotznik, Ph.D. Professor The University of Delaware Tel: (302) EEG 05 Fall 07 ecure #0 Mark Mirznik, Ph.D. Prfessr The Universiy f Delaware Tel: (3083-4 Email: mirzni@ece.udel.edu haper 7: apacirs and Inducrs The apacir Symbl Wha hey really lk like The apacir Wha

More information

3 ) = 10(1-3t)e -3t A

3 ) = 10(1-3t)e -3t A haper 6, Sluin. d i ( e 6 e ) 0( - )e - A p i 0(-)e - e - 0( - )e -6 W haper 6, Sluin. w w (40)(80 (40)(0) ) ( ) w w w 0 0 80 60 kw haper 6, Sluin. i d 80 60 40x0 480 ma haper 6, Sluin 4. i (0) 6sin 4-0.7

More information

Christos Papadimitriou & Luca Trevisan November 22, 2016

Christos Papadimitriou & Luca Trevisan November 22, 2016 U.C. Bereley CS170: Algorihms Handou LN-11-22 Chrisos Papadimiriou & Luca Trevisan November 22, 2016 Sreaming algorihms In his lecure and he nex one we sudy memory-efficien algorihms ha process a sream

More information

CHAPTER 14 CHEMICAL KINETICS

CHAPTER 14 CHEMICAL KINETICS CHAPTER 4 CHEMICAL KINETICS PRACTICE EXAMPLES A B (E) The rae f cnsumpin fr a reacan is expressed as he negaive f he change in mlariy divided by he ime inerval. The rae f reacin is expressed as he rae

More information

Notes for Lecture 17-18

Notes for Lecture 17-18 U.C. Berkeley CS278: Compuaional Complexiy Handou N7-8 Professor Luca Trevisan April 3-8, 2008 Noes for Lecure 7-8 In hese wo lecures we prove he firs half of he PCP Theorem, he Amplificaion Lemma, up

More information

Strengthening of web opening in non-compact steel girders

Strengthening of web opening in non-compact steel girders IOSR Jurnal f Mechanical and Civil Engineering (IOSR-JMCE) e-issn: 2278-1684,p-ISSN: 2320-334X, Vlume 12, Issue 5 Ver. II (Sep. - Oc. 2015), PP 34-47 www.isrjurnals.rg Srenghening f web pening in nn-cmpac

More information

PASSIVE PFC FOR FLYBACK CONVERTORS

PASSIVE PFC FOR FLYBACK CONVERTORS PASSIE PFC FOR FLYBACK COERORS Parviz Par and Keyue M Sedley Dep f Elecrical and Cpuer Engineering Universiy f Califrnia, Irvine Irvine, Califrnia 9697 Absrac A new passive Pwer Facr Crrecr (PFC) based

More information

Nature Neuroscience: doi: /nn Supplementary Figure 1. Spike-count autocorrelations in time.

Nature Neuroscience: doi: /nn Supplementary Figure 1. Spike-count autocorrelations in time. Supplemenary Figure 1 Spike-coun auocorrelaions in ime. Normalized auocorrelaion marices are shown for each area in a daase. The marix shows he mean correlaion of he spike coun in each ime bin wih he spike

More information

PENALIZED LEAST SQUARES AND PENALIZED LIKELIHOOD

PENALIZED LEAST SQUARES AND PENALIZED LIKELIHOOD PENALIZED LEAST SQUARES AND PENALIZED LIKELIHOOD HAN XIAO 1. Penalized Leas Squares Lasso solves he following opimizaion problem, ˆβ lasso = arg max β R p+1 1 N y i β 0 N x ij β j β j (1.1) for some 0.

More information

1 Review of Zero-Sum Games

1 Review of Zero-Sum Games COS 5: heoreical Machine Learning Lecurer: Rob Schapire Lecure #23 Scribe: Eugene Brevdo April 30, 2008 Review of Zero-Sum Games Las ime we inroduced a mahemaical model for wo player zero-sum games. Any

More information

E β t log (C t ) + M t M t 1. = Y t + B t 1 P t. B t 0 (3) v t = P tc t M t Question 1. Find the FOC s for an optimum in the agent s problem.

E β t log (C t ) + M t M t 1. = Y t + B t 1 P t. B t 0 (3) v t = P tc t M t Question 1. Find the FOC s for an optimum in the agent s problem. Noes, M. Krause.. Problem Se 9: Exercise on FTPL Same model as in paper and lecure, only ha one-period govenmen bonds are replaced by consols, which are bonds ha pay one dollar forever. I has curren marke

More information