Neural Networks Trained with the EEM Algorithm: Tuning the Smoothing Parameter

Size: px
Start display at page:

Download "Neural Networks Trained with the EEM Algorithm: Tuning the Smoothing Parameter"

Transcription

1 eural etworks Trained wit te EEM Algorit: Tuning te Sooting Paraeter JORGE M. SATOS,2, JOAQUIM MARQUES DE SÁ AD LUÍS A. ALEXADRE 3 Intituto de Engenaria Bioédica, Porto, Portugal 2 Instituto Superior de Engenaria do Porto, Porto, Portugal 3 IT etworks and Multiedia Group, Covilã, Portugal js@isep.ipp.pt, jsa@fe.up.pt and lfbaa@di.ubi.pt Abstract: - Te training of eural etworks and particularly Multi-Layer Perceptrons (MLP's) is ade by iniizing an error function usually known as "cost function". In our previous works, we apply te Error Entropy Miniization (EEM) algorit in classification and its optiized version using, as cost function, te entropy of te errors between te outputs and te desired targets of te neural network. One of te difficulties in ipleenting te EEM algorit is te coice of te sooting paraeter, also known as window size, in te Parzen Window probability density function estiation for te coputation of te entropy and its gradient. We present ere a forula yielding te value of te sooting paraeter, depending on te nuber of data saples and on te neural network output diension. Several experients wit real data sets were ade in order to sow te validity of te proposed forula. Key-Words: - Entropy, Parzen, Sooting Paraeter, Cost Function Introduction Te use of entropy and related concepts in learning systes is well known. Te Error Entropy Miniization concept was introduced in [] and we use te sae approac in [2] wit te goal of perforing classification. Te EEM Algorit for classification use, as cost function, te Renyi's Quadratic Entropy of te error, between te output of te neural network and te desired targets. We ave ade a furter optiization of te EEM algorit in [3] introducing a variable learning rate to acieve bot a faster convergence of te algorit and an autoatic adaptation of te value of te learning rate to te specific proble. In order to copute te Renyi's Quadratic Entropy of te error one ust coose a value for te sooting paraeter in te Parzen Window etod, tat is best suited for a specific data set. In all known works using tis etod, giving as exaple ours in [2,3] and oter autors in [,4], te value of te sooting paraeter is always experientally selected. Knowing tat tis approac is not te ost appropriate, in tis paper, we study te beavior of te algorit wit different data sets and propose a forula to copute te sooting paraeter for a specific data set and neural network. In section 2 te EEM algorit and furter optiizations are presented. In section 3 te sooting paraeter for entropy and gradient coputation is discussed and a specific forula is presented. In section 4 te results of several experients are presented sowing te validity of te forula. In te final section, we present te conclusions. 2 Te EEM Algorit Let y = { y i } R, i =,..,, be a set of saples fro te output vector Y R of a apping n R R : Y = g( w, X ), were w is a set of eural etwork () weigts, X is te input vector and is te diensionality of te output vector. Let d = d i {, } be te desired targets and ei = d i yi te error for eac data saple. In order to copute te Renyi's Quadratic Entropy of e we use te Parzen Window probability density function (pdf) estiation. Tis etod estiates te pdf as e ei f ( e) = K( ) () i= were K is a kernel function and te bandwidt or sooting paraeter. We use te Gaussian kernel wit zero ean and unitary covariance atrix: = T G( e, I) exp ( e e ) 2 (2π ) 2 (2) Te Reniy's Quadratic Entropy of te error e can be estiated, applying te integration of gaussian kernels [5], by Hˆ ( e) log 2 = logv ( e) = 2 ei e j G(,2I) i= j= (3)

2 Te gradient of V (e) is: ei e j Fi = G(,2I )( e ) 2 i e + j (4) 2 j= Tis gradient is back-propagated into te MLP te sae way as in te MSE algorit. Te update of te neural network weigts is perfored using w = ± η V. Te ± eans tat we can iniize w ( ) or axiize (+) te entropy. We also proved in [2] tat we can use tis cost function for classification if te output of te neural network is on [ a, a] and te targets are in { a, a}, a R. Since one of te probles of te EEM algorit is its slow convergence, we introduced in [3] an optiization of te algorit based on te use of a variable learning rate (VLR), during te training pase, wic applies te rule: ( n ) ( n) ( n ) η u if H < ( ) H n η = (5) ( n ) ( n) ( n ) η d if H H being u and d te increasing (up) and decreasing (down) factors of te learning rate η in consecutive iterations. Tis optiization iproved te perforance of te original algorit and we called it te EEM-VLR algorit. 3 Sooting Paraeter In te EEM algorit, te error entropy estiation is different according to te vector e diension. Tis diension depends on te nuber of MLP output neurons tat depends on te nuber of classes, C, and teir coding. If we use binary coding, te diension of vector e will be log 2 C wereas if we use te -out-of- C coding te nuber of output neurons as well as te diension of vector e will be equal to te nuber of classes. In te experients tat we perfored, we obtained good results for bot codings but we suggest te use of binary coding only wen te nuber of classes is a power of 2; oterwise, we get an excessive nuber of outputs copared wit te nuber of classes. One of te probles of pdf estiation using te Parzen Window etod, besides te coice of te kernel, is te coice of te sooting paraeter. In te EEM algorit te value of depends on: te different coding of te nuber of classes; te nuber of data saples; te diension of te vector e. For continuous f (x) te estiated density function will converge to te true density as, wen: and (6) Silveran proposed in [6], for unidiensional cases assuing Gaussian distributed saples, and using a Gaussian kernel, te following forula for te sooting paraeter:.2 op =.6σ (7) were σ is te saple standard deviation and is te nuber of saples. For ultidiensional cases, also assuing noral distributions and using te noral kernel, Bowan and Azzalini [7] proposed te forula: op = σ ( 2) + (8) were is te diension of vector x. An iportant fact tat ipedes, in our case, te use of forulas 7 and 8 is tat our algorit uses te entropy of e as a control variable, i.e., te algorit progresses only if te entropy at a given iteration is saller tan at te previous one. Since te entropy value is proportional to te sooting paraeter value, used to copute it, if one uses a value for proportional to te variance of e, one igt be increasing, by tis siple fact, te entropy value and te algorit fails to reac te iniu. Tis eans tat we are liited to use a value for tat does not depend on σ. Considering tat variable e takes values, in te unidiensional case, in te interval [ 2,2], and tat te axiu standard deviation in tis case is 2, we considered to use tis value to replace te standard deviation in forulas 7 and 8. So, tese forulas becae: and.2 op = 2.2 (9) op = 2 ( 2) + () ote tat, in te EEM algorit, we only need to copute te entropy and its gradient; we do not

3 need to estiate te probability density function of e. Tis is a relevant fact because, in te gradient descent etod, ore iportant tan coputing wit extree precision te gradient is to get wit relative precision its direction. In addition, te coputation wit extree accuracy of te probability density function causes te entropy to ave ig variability. Tis fact could lead to te occurrence of local inia. Given tese considerations, we do not esitate in using values iger tan te ones usually proposed for pdf estiation. Taking into account te experiental results wit several data sets we tried to forulate a rule tat yields iger values of for saller data sets tan tose obtained wit forula and still inducing te sae beavior. We ten arrived at te following proposed new forula wit siilar beavior as () op = 25 () otice te decreasing beavior wit and increasing beavior wit as sould be desired. A coparison between te values of obtained wit forulas and for different values of is sown in Fig.. In te several experients tat we perfored using forula te results were very satisfactory, as we will see in te next section. 4 Experients We perfored several experients wit different real data sets, wit different nuber of saples and different nuber of classes, suarized in Table Table Real data sets used for te experients. Data sets Saples Features Classes Ionospere Sonar wdbc Iris Wine Pb (Te Pb2 data set can be found in [8] and all te oters can be found in te UCI repository of acine learning database). We also produced an artificial data set and used it as a 2-class and as a 4- class classification proble to try to establis te influence of te nuber of classes in te value of te sooting paraeter. Te two versions of te artificial data set (tat we call XOR-n), siilar to an XOR proble, but wit soe noise added, are sown in Fig = = = Fig. - Value of for forulas (dased) and (solid) for =2, 3 and 4. (Marked points refer to experients wit data sets entioned in section 4).

4 In all experients we used (I,n,) MLP's, were I is te nuber of input neurons, n is te nuber of neurons in te idden layer and is te nuber of output neurons. We applied te cross-validation etod using alf of te data for training and alf for testing. Te experients for eac data set were perfored varying te nuber of neurons in te idden layer, te value of te sooting paraeter and using different nuber of epocs XOR n 2 classes XOR n 4 classes.5.5 Fig. 2 - Artificial data set for te 2 first experients. Te results are sown in Table 2 and Table 4. Eac result is te ean error of 2 repetitions. For eac experient we igligted te best results perfored wit te sae nuber of epocs in order to get te needed guidance about te optiu value for te sooting paraeter. In Table 2 we sow te results of te classification errors for te XOR-n data sets. Table 2 Errors (%) for XOR-n data sets, 2 and 4 classes (Resp. 8, 2 and 6 epocs fro top to botto) XOR-n 2 classes uber of Hiden-Layer eurons , 36,5 28,25 24,4 23, 23,38 27,3 2,4 35,43 3,38 26,43 23,93 27, 28,5 2,8 37,9 27,8 24,45 26,65 26,38 26,48 3,2 36,28 26,33 26,78 23,88 25,65 23,68 3,6 35,93 26,3 27, 22,43 27,8 27,2 4, 35,4 25,75 26,83 24,93 2, 24,8 4,4 35,33 27,75 27,25 22,3 26,65 22,2 4,8 35,63 24,7 24,98 23,73 25,33 22,65 5,2 34,45 28,6 26,75 23,5 2,2 24,68 5,6 36,35 25,68 23,75 24,93 25,5 24, 6, 34,2 27,53 23,7 22,38 25,55 23,6 2, 35, 24,3 24,43 22,75 24,48 23,55 2,4 35,78 29,68 25,8 23,88 25,75 26,38 2,8 34, 26,25 25, 25,45 26,75 27,3 3,2 33,68 25,5 24,8 28,55 26,63 25,63 3,6 33,83 24,65 24,53 25,3 24,7 26,43 4, 34,95 25,8 2,2 22,88 26,23 22,5 4,4 33,2 23,38 24, 25,8 23,3 24,5 4,8 33,98 24, 23,68 23,88 23,43 23, 5,2 32,93 22,6 23,63 24,8 24,2 25,8 5,6 33,78 23,4 22,75 23,8 24,93 23,25 6, 33,98 25,8 23,33 22,75 22,63 24,5 2, 36,28 27,3 22,78 23,43 23,7 22,55 2,4 34,68 26,28 24,93 24,3 26,6 27,5 2,8 35, 27, 24,4 27,43 25,5 26,55 3,2 33,65 26,58 24,75 25,25 26,6 25,83 3,6 34,9 22, 26,48 24,8 24,98 25, 4, 33,78 25,63 2,28 24,33 25,8 23,48 4,4 34,3 24,4 23,58 25,53 23,73 26,73 4,8 33,68 24, 23,38 22,55 22,78 23,7 5,2 35, 2,9 24,45 23,25 23,38 23, 5,6 33,3 2,33 23,8 24,33 23,6 23,6 6, 34,7 22,3 22,8 24,33 22,95 23,33 XOR-n 4 classes uber of Hidden-Layer eurons , 9, 9,8 8,3 9,3 8,88 9, 2,4 7,65 8,3 9, 8,73 7,4 8,83 2,8 8,9 7,98 7,9 7,85 9,5 8,23 3,2 7,75 8,8 7,73 8,3 7,75 8,58 3,6 9, 8,5 7,95 8,3 8,53 8,3 4, 7,38 7,73 8,65 8,2 9,3 8,38 4,4 8,8 8,78 8, 8,3 8,8 8,7 4,8 8,68 7,98 8,48 9,8 8,5 8,25 5,2 8,83 7,7 7,3 9,53 8,7 8,7 5,6 8,68 8,33 7,65 8,4 9,5 9,3 6, 7,4 9, 8, 9,53 7,73 8,45 2, 8,63 9,43 8,63 2,8 9,7 2,63 2,4 2,38 9,28 8,55 8,95 9,73 2,45 2,8 8,33 8,98 9,3 8,48 8,88 8,38 3,2 8,28 8,23 7,75 8,5 8,53 8,65 3,6 8,2 8,5 8,95 8,73 7,75 8,8 4, 8,48 8,23 8,8 8,7 8,5 7,78 4,4 7,68 8,8 8,98 8,48 8,5 9,5 4,8 7,6 8,6 8,93 8,35 8,8 8,5 5,2 7,38 8,8 7,8 7,95 8,63 8,78 5,6 7,5 8,23 7,68 8,65 8,78 9,5 6, 7,85 8,8 7,85 8,8 7,68 8,55 2, 8,35 9,45 9,8 9,2 8,83 9,48 2,4 8,6 8,63 9, 9,3 9,85 9,8 2,8 7,93 9,5 9,75 9,6 9,35 9,38 3,2 8,88 9,38 9,68 9,35 9,38 9,58 3,6 8,8 7,95 8,9 8,9 9,9 9,5 4, 7,33 8,85 9,3 8,28 9,65 9,28 4,4 8,43 2,8 9,28 9,6 9,45 9,53 4,8 8,38 9,3 9,3 9,5 8,98 9,8 5,2 7,4 9,68 8,85 8,78 9,98 9,33 5,6 7,6 8,45 8,33 9,3 9,58 9,8 6, 8,8 8,7 9,45 9,3 9,38 9,8

5 Coparing te two tables, we can see tat, an increased nuber of classes deand an increased value of te sooting paraeter. In te first case, (2-class proble), te optiu value for is about 4. and in te second case, (4-class proble), te optiu value for is about 4.8. Te classification errors for te real data sets are sown in Table 4. For eac data set, we perfored experients wit different nuber of epocs. In Table 3 we present te values of for te iniu classification errors ( Best colun) and te value of tat represents, in our opinion, tose sallest classification errors ( Suggested colun). We also present te values proposed for eac data set by forulas and. Data sets Table 3 Values of for eac data set driven by te experients and defined by forulas and. r. of Classes r. of Saples Best Suggested ( in. Errors) For. For. Ionospere ,6 4,2 2,67,85 Sonar ,2 3,9 3,47,92 wdbc 2 569,4,6 2,,78 Xor , 4,6 3,54,93 Iris 3 5 4, 3,8 5,,5 Wine ,6 4,2 4,59,2 Pb2 4 68,8 2, 2,87,93 Xor ,2 4,8 5,,7 Te experients clearly sow tat, for sall values of, te classification errors are significantly large and, terefore, te use of forulas like (), used for density estiation, are not appropriate for te EEM algorit. Figure sows (crosses and open circles), te values obtained in te experients wit all data sets were we can easily see teir relation wit te values of te proposed forula. For eac data set te crosses are te suggested values given by te sallest classification errors and te circles are te values for te iniu classification error. 5 Conclusions In tis paper, we ave presented a forula for te value of te bandwidt paraeter,, of te entropy estiation, to use in te EEM-VLR algorit. Tis forula, contrary to te ones proposed by Silveran and Bowan, does not depend on σ due to te iterative nature of te EEM-VLR algorit. We sowed tat it produces very good results using a set of experients were te values proposed by our forula are uc closer to te best ones found epirically, tan te values obtained using te Silveran and Bowan forulas. We expect tat tis contribution ay save tie and effort to any practitioners using tis type of iterative entropic algorits involving Parzen Window density estiation. Acknowledgent Tis work was supported by te Portuguese Fundação para a Ciência e Tecnologia (project POSI/EIA/5698/24). References: [] D. Erdogus and J. Principe, An Error-Entropy Miniization Algorit for Supervised Training of onlinear Adaptive Systes, Trans. On Signal Processing, Vol. 5, o. 7, 22, pp [2] J. M. Santos, L. A. Alexandre, and J. Marques de Sá, Te error entropy iniization algorit for neural network classification, in Proceedings of te 5t International Conference on Recent Advances in Soft Coputing, A. Lofti, Ed. ottinga Trent University, 24, pp [3] J. M. Santos, J. Marques de Sá, L. A. Alexandre, and F. Sereno, Optiization of te error entropy iniization algorit for neural network classification, in Intelligent Engineering Systes Troug Artificial eural etworks, C.H.Dagli, A. L. Buczak, D. L. Enke, M. J. Ebrects, and O. Ersoy, Eds., vol. 4. ASME Press, 24, pp [4] D. Erdogus and J. Principe, "Entropy Miniization Algorit for eural etworks," Proceedings of te Intl. Joint Conf. on eural etworks, 2, pp [5] D. Xu and J. Princpe, Training lps layer-bylayer wit te inforation potential, in Intl. Joint Conf. on eural etworks, 999, pp [6] B. W. Silveran, Density Estiation for Statistics and Data Analisys. Capan & Hall, 986, vol. 26. [7] A. W. Bowan and A. Azzalini, Applied Sooting Tecniques for Data Analysis. London: Oxford University Press, 997. [8] R. Jacobs, M. Jordan, S. owlan, and G. Hinton, Adaptive ixtures of local experts, eural Coputation, pp. pp.79 87, 99.

6 Table 4 - Errors (%) for te data sets Ionospere (resp. 4, 6 and 8 epocs), Sonar (resp. 5, and 5 epocs),wdbc (resp. 4, 6 and 8 Epocs), Iris (resp. 4, 6 and 8 epocs),wine (resp. 4, 6 and 8 epocs) and Pb2 (resp. 2, 25 and 3 epocs) for different nuber of idden-layer neurons and several values of te sooting paraeter. Ionospere - 4, 6 and 8 epocs ,4 33,84 9,54 2,4 5,29 6,26 7,3 2,7 2,76 2,97 3,24 3,54 3,27 3,7 3,2 3, 3,9 3,7 3,49,8 2,47 2,7 3,2 3,3 2,69 2,6 3,3 3,9 2,9 3, 2,8 3,4 3,47 3,28 3,4 3,8 2,87 3,2 2,2 2,7 2,86 2,8 3,2 3,3 3, 3,4 3,6 2,88 2,9 2,8 2,83 4,2 3,4 3,4 3, 2,8 3,5 2,6 3,7 2,4 2,37 2,82 2,93 3,7 3,7 2,83 3,6 2,94 3,6 2,7 3,33 2,83 2,94 2,53 3,6 2,69 3 2,7 2,97 2,74 2,77 2,7 3,23 3,7 3,3 3,23 2,86 2,84 3,9 3, 2,7 2,7 2,33 2,9 2,9 3,4 3, 2,8 2,8 2,67 2,64 2,84 2,9 2,5 2,77 2,66 2,94 2,6 3, 3, 3, 2,73 2,83 2,59 3,8 2,73 2,79 2,7 2,26 2,46 2,26 3,7 3, 3,36 2,89 2,24 2,87 2,8 2,62 2,88 2,24 2,69 2,63 4,2 3,9 2,39 2,67 2,73 2,59 2,97 3,6 3,29 2,82 2,46 2,4 2,4 3,4 2,4 2,64 2,34 2,94 2,99 4,6 2,96 2,8 2,6 2,8 2,43 2,63 2,77 2,7 2,6 2,44 2,29 2,96 2,6 2,4 3,29 2,67 2,73 2,72 5 2,7 2,6 2,9 2,7 2,69 2,87 3, 2,44 2,44 2,29 3,6 2,87 3,6 2,56 2,6 2,6 2,63 3,6 Sonar - 5, and 5 epocs , 49,47 48,63 49,8 48,97 48,56 5,65 48,36 49,9 49,3 49,4 48,4 49,8 5,82 5,48 5, 49,86 5,82 49,95,3 48,49 45,77 45,67 43,22 4,37 43,73 36,7 35,36 37,74 35,53 34,52 3,42 3,63 28, 27,8 29,88 26,9 25,5,6 24,9 23,65 23,6 22,96 22,88 23,7 24,64 23,44 23,4 24,6 23,94 23,8 25,53 24,52 23,92 24,8 25, 23,6,9 24,78 23,27 22,96 22,4 22,65 23,3 24,2 24,52 23,5 22,86 23,29 23,8 24,23 24,52 23,8 22,33 24,3 24,52 2,2 23,7 23,68 23,53 24,28 23,77 23,66 24,6 24,33 23,49 22,84 24,42 23,8 25,3 24,45 23,27 22,67 22,74 23,5 2,5 24,37 23,5 22,28 23,8 23,34 24,8 23,99 22,4 24,52 24,35 22,8 22,74 24,6 24,6 22,26 22,67 23,4 23,25 2,8 23,82 23,7 23,6 22,72 22,76 22,36 23,5 24,47 23, 22,8 23, 23,87 24,33 23,85 23,99 23, 22,86 22,67 3, 24,4 24,6 22,98 22,98 22,98 22,8 24,64 24,4 22,45 23,82 23,29 22,74 24,98 23,7 23,29 23,5 22,26 22,88 3,4 24, 23,82 24,8 24,4 23,8 23,3 24,6 24,35 23,5 22,2 23,27 2,32 24,42 23,56 22,4 23,32 23,7 22, 3,7 24,5 23,82 2,73 22,8 22,2 22,8 23,56 23,58 23,37 23,44 23,44 2,6 25,4 23,58 23,5 22,84 23,27 22,7 4, 24,26 23,5 23,42 23,56 22,2 22,96 23,58 23,77 23,3 22,4 22,67 22,4 23,73 23,89 23,8 2,95 22,7 22,6 Wisconsin breast cancer - 4, 6 and 8 epocs , 5,34 52,3 5,22 49,44 46,88 5,54 5,6 49,32 47,58 45,97 48,8 46,52 49,48 48,6 48,58,2 8,83,37 9,4 9,74 2,8,7,46,48 9,74 9,47 4,4 2,89 5,49 5,83 6,,4 2,53 2,33 2,65 2,57 3, 2,79 2,8 3,7 2,95 2,82 2,77 2,8 3,8 3, 2,92,6 2,77 2,7 2,6 2,7 2,77 2,97 2,83 2,9 2,83 2,96 2,95 3,23 3,75 3,6 2,9,8 2,83 2,75 2,9 2,84 2,78 3,5 3,6 3,6 2,97 3,5 2,96 3,4 3, 3,8 3,8 2, 2,84 2,87 2,64 2,58 2,97 2,97 3, 3,6 3,7 3,8 3,26 2,87 3,2 2,99 3,6 2,2 2,9 2,77 2,9 3,7 2,9 2,72 3,6 2,8 2,97 3,23 3,3 3, 3,7 2,95 3,5 2,4 2,67 2,76 2,92 2,84 2,59 2,94 3,4 3,4 2,88 3,24 3,6 3,26 3,22 2,99 3,36 2,6 3,28 2,88 3,3 2,97 3,28 3,7 2,93 3,2 3,6 3,49 3,3 3,8 3,6 3,2 3,34 2,8 2,89 2,85 2,79 2,94 2,95 3,6 3,2 3,25 3,7 3,28 3,5 3,5 3,28 3,2 3,4 3, 2,69 2,9 3,9 3,9 2,98 2,98 3,4 3,6 3,7 3,37 3,3 3,49 3,49 3,35 3,35 Iris - 4, 6 and 8 epocs , 7,97 7,2 6,33 9,2 7,27 5,7 5,23 5,7 4,4 4,9 5,3 4,57 4, 4,23 4,3 2,4 7,7 6,63 4,6 4,63 4,24 4,8 3,67 4,5 3,87 4,3 7,83 3,73 4,8 4,23 4,23 2,8 6,77 4,43 4,77 3,83 4,7 5,87 4,67 4,27 4,7 4,3 4,6 4,5 3,77 3,57 4, 3,2 7,37 4,87 4, 4,4 4,4 8,4 4, 3,93 4,27 3,8 5,27 4,4 3,73 3,83 3,93 3,6 4,7 6,67 4,67 4, 4,3 6,37 5, 3,97 3,97 4,3 4,57 4,7 4,63 4, 4, 4, 5,5 4,3 4,83 5,5 4,53 6,6 3,97 3,97 3,87 4,53 5,57 5,73 3,83 3,5 3,8 4,4 4,8 4,2 4,37 4,73 4,23 6,6 4,47 3,8 4,2 4,37 7,3 4,23 3,67 4,3 4,57 4,8 5,5 4, 4,33 4,57 4,37 6,37 3,8 4,4 4,5 4,7 5,7 4,53 4,3 4,2 4, 5,2 7,4 4,67 4,63 4,97 4,5 4,6 4,23 4,7 4,3 4, 6, 4,53 4,53 3,9 3,87 5,6 5,54 3,9 4,2 4,47 4,27 5,4 4,53 4,2 4,4 4, 4,7 4,3 4,57 4, 3,73 6, 5, 4,37 4,7 4, 4,53 5,7 4,3 4,43 4,3 4, 5, 5, 4,3 3,83 3,8 Wine - 4, 6 and 8 epocs ,4 46,4 34,92 29,3 22,89 7,89 8,6 7,75 6,77 2,25 3,79 2,3 2,8 2,6 2,56 2,3 2,92 2,67 2,56 2,39 2,67 2,39,8 8,54 2,6 2,28 2,28 2,28 2,53 2,45 2,6 2,64 2,33 2,6 2,39 2,5 2,36 2,47 2,64 2,42 2,75 2,64 2,95 2,59 2,2 2,28 2,44 2,9 2,56 2,5 2,67 2,36 2,95 2,28 2,42 2,56 2,6 2,59 2,3 2,53 2,64 2,6 2,73 2,59 2,28 2,64 2,6 2,3 2,84 2,25,94 2,36 2,42 2,33 2,28 2,39 2,9 2,4 2,36 2,33 2,28 2,47 2,8 2,64 2,78 2,47 2,64 2,89 3, 2,47 2,42 2,36,97 2,39 2,3 2,22 2,42 2,39 2,44 2,8 2,56 2,28 2,22 2,8 2,56 2,42 2,84 2,6 2,87 2,3 3,4 2,33 2,59 2,9 2,5 2,78 2,8,94 2,5 2,22 2,89 2,22 2,3 2,6 2,45 2,7 2,33 2,44 2,42 2,25 2,47 2,42 3,8 2,64 2,59 2,4 2, 2,3 2,3 2,33 2,5 2,39 2,64,94 2,42 2,64 2,22 2,8 2,59 2,84 2,7 2,78 2,36 2,75 4,2 2,4 2,28 2,22 2, 2,64,94 2,42 2,45 2,36,88 2,75 2,45 2,9 2,44 2,78,88 2,67 2,36 2,87 2,3 2,53 4,6 2,56 2,28 2,6 2,39 2,,83,9 2,47 2,3 2,78 2,39 2,6 2,28 2,6 3, 2,36 2,5 2,95 2,39 2,89 2,87 5, 2,5,86 2,42 2,6 2,22,97 2,2 2,89 2,53 2, 2,36 2,7 2,39 2,28 3,6 2,84 2,6 2,44 2,2 2,42 2,7 Pb2-2, 25 and 3 epocs ,6 24,9 6,7 3,26 4,46 4,7 4,62 22,29 5,7 2,76,2,65,23 8,73 2,,63,78,68 9,3,8 2,25,53 8,9 9,29 8,88 8,2 22,89 7,8 8,5 8,29 7,76 8,36 25,8 8, 7,5 7,89 8,39 8,25 2, 25,49 9, 7,82 8, 8,5 8,92 24,94 7,65 7,82 8,4 7,64 9,8 25,7 8,36 8,43 7,55 8,8 8,7 2,2 28,72 8,79 8,66 8,49 8,85 9,3 28,46,35 8,24 7,8 7,58 8,92 27,2 9,4 8,7 7,94 8,5 7,6 2,4 28,64,65 8,77 9,8 8,35 9,88 3,83 2,38 9,6 8,53 8,2 8,8 29,65,77 8,4 8,7 9,76 7,76 2,6 28,94 9,73 9,9 9,34 9,7 8,73 3,3 3,22 8,77,58 9,8 9,26 29,3,78 9,43 7,72 8,3 8,9 2,8 28,77,99,28 9,36 9,6 9,36 28,75,97 8,9 8,22 9,89 9,38 28,63,2 8,77 9,83 9,37 8,82 3, 29,35,5,37 9,5 8,3 9,89 28,29,8 9,6,43,5 8,63 28,3,55 7,55 8,26 8,5 8,89

Derivative at a point

Derivative at a point Roberto s Notes on Differential Calculus Capter : Definition of derivative Section Derivative at a point Wat you need to know already: Te concept of liit and basic etods for coputing liits. Wat you can

More information

Intelligent Systems: Reasoning and Recognition. Perceptrons and Support Vector Machines

Intelligent Systems: Reasoning and Recognition. Perceptrons and Support Vector Machines Intelligent Systes: Reasoning and Recognition Jaes L. Crowley osig 1 Winter Seester 2018 Lesson 6 27 February 2018 Outline Perceptrons and Support Vector achines Notation...2 Linear odels...3 Lines, Planes

More information

lecture 35: Linear Multistep Mehods: Truncation Error

lecture 35: Linear Multistep Mehods: Truncation Error 88 lecture 5: Linear Multistep Meods: Truncation Error 5.5 Linear ultistep etods One-step etods construct an approxiate solution x k+ x(t k+ ) using only one previous approxiation, x k. Tis approac enoys

More information

Intelligent Systems: Reasoning and Recognition. Artificial Neural Networks

Intelligent Systems: Reasoning and Recognition. Artificial Neural Networks Intelligent Systes: Reasoning and Recognition Jaes L. Crowley MOSIG M1 Winter Seester 2018 Lesson 7 1 March 2018 Outline Artificial Neural Networks Notation...2 Introduction...3 Key Equations... 3 Artificial

More information

1 Proving the Fundamental Theorem of Statistical Learning

1 Proving the Fundamental Theorem of Statistical Learning THEORETICAL MACHINE LEARNING COS 5 LECTURE #7 APRIL 5, 6 LECTURER: ELAD HAZAN NAME: FERMI MA ANDDANIEL SUO oving te Fundaental Teore of Statistical Learning In tis section, we prove te following: Teore.

More information

Pattern Recognition and Machine Learning. Artificial Neural networks

Pattern Recognition and Machine Learning. Artificial Neural networks Pattern Recognition and Machine Learning Jaes L. Crowley ENSIMAG 3 - MMIS Fall Seester 2017 Lessons 7 20 Dec 2017 Outline Artificial Neural networks Notation...2 Introduction...3 Key Equations... 3 Artificial

More information

c hc h c h. Chapter Since E n L 2 in Eq. 39-4, we see that if L is doubled, then E 1 becomes (2.6 ev)(2) 2 = 0.65 ev.

c hc h c h. Chapter Since E n L 2 in Eq. 39-4, we see that if L is doubled, then E 1 becomes (2.6 ev)(2) 2 = 0.65 ev. Capter 39 Since n L in q 39-4, we see tat if L is doubled, ten becoes (6 ev)() = 065 ev We first note tat since = 666 0 34 J s and c = 998 0 8 /s, 34 8 c6 66 0 J sc 998 0 / s c 40eV n 9 9 60 0 J / ev 0

More information

Estimation for the Parameters of the Exponentiated Exponential Distribution Using a Median Ranked Set Sampling

Estimation for the Parameters of the Exponentiated Exponential Distribution Using a Median Ranked Set Sampling Journal of Modern Applied Statistical Metods Volue 14 Issue 1 Article 19 5-1-015 Estiation for te Paraeters of te Exponentiated Exponential Distribution Using a Median Ranked Set Sapling Monjed H. Sau

More information

Block designs and statistics

Block designs and statistics Bloc designs and statistics Notes for Math 447 May 3, 2011 The ain paraeters of a bloc design are nuber of varieties v, bloc size, nuber of blocs b. A design is built on a set of v eleents. Each eleent

More information

Ch 12: Variations on Backpropagation

Ch 12: Variations on Backpropagation Ch 2: Variations on Backpropagation The basic backpropagation algorith is too slow for ost practical applications. It ay take days or weeks of coputer tie. We deonstrate why the backpropagation algorith

More information

Pattern Recognition and Machine Learning. Artificial Neural networks

Pattern Recognition and Machine Learning. Artificial Neural networks Pattern Recognition and Machine Learning Jaes L. Crowley ENSIMAG 3 - MMIS Fall Seester 2016 Lessons 7 14 Dec 2016 Outline Artificial Neural networks Notation...2 1. Introduction...3... 3 The Artificial

More information

Determining Limits of Thermal NDT of Thick Graphite/Epoxy Composites

Determining Limits of Thermal NDT of Thick Graphite/Epoxy Composites ECNDT 006 - We.3.8.1 Deterining Liits of Teral NDT of Tick Grapite/Epoy Coposites Vladiir VAVILOV Institute of Introscopy Tosk Russia Abstract. Te known approac to inspecting tin coposites by using infrared

More information

CS Lecture 13. More Maximum Likelihood

CS Lecture 13. More Maximum Likelihood CS 6347 Lecture 13 More Maxiu Likelihood Recap Last tie: Introduction to axiu likelihood estiation MLE for Bayesian networks Optial CPTs correspond to epirical counts Today: MLE for CRFs 2 Maxiu Likelihood

More information

Supplementary Materials: Proofs and Technical Details for Parsimonious Tensor Response Regression Lexin Li and Xin Zhang

Supplementary Materials: Proofs and Technical Details for Parsimonious Tensor Response Regression Lexin Li and Xin Zhang Suppleentary Materials: Proofs and Tecnical Details for Parsionious Tensor Response Regression Lexin Li and Xin Zang A Soe preliinary results We will apply te following two results repeatedly. For a positive

More information

Estimating the Density of a Conditional Expectation

Estimating the Density of a Conditional Expectation Estiating te Density of a Conditional Expectation Sauel G. Steckley Sane G. Henderson David Ruppert Ran Yang Daniel W. Apley Jerey Stau Abstract In tis paper, we analyze etods for estiating te density

More information

Kernel Methods and Support Vector Machines

Kernel Methods and Support Vector Machines Intelligent Systes: Reasoning and Recognition Jaes L. Crowley ENSIAG 2 / osig 1 Second Seester 2012/2013 Lesson 20 2 ay 2013 Kernel ethods and Support Vector achines Contents Kernel Functions...2 Quadratic

More information

Submanifold density estimation

Submanifold density estimation Subanifold density estiation Arkadas Ozakin Georgia Tec Researc Institute Georgia Insitute of Tecnology arkadas.ozakin@gtri.gatec.edu Alexander Gray College of Coputing Georgia Institute of Tecnology agray@cc.gatec.edu

More information

LAB #3: ELECTROSTATIC FIELD COMPUTATION

LAB #3: ELECTROSTATIC FIELD COMPUTATION ECE 306 Revised: 1-6-00 LAB #3: ELECTROSTATIC FIELD COMPUTATION Purpose During tis lab you will investigate te ways in wic te electrostatic field can be teoretically predicted. Bot analytic and nuerical

More information

Sharp Time Data Tradeoffs for Linear Inverse Problems

Sharp Time Data Tradeoffs for Linear Inverse Problems Sharp Tie Data Tradeoffs for Linear Inverse Probles Saet Oyak Benjain Recht Mahdi Soltanolkotabi January 016 Abstract In this paper we characterize sharp tie-data tradeoffs for optiization probles used

More information

Non-Parametric Non-Line-of-Sight Identification 1

Non-Parametric Non-Line-of-Sight Identification 1 Non-Paraetric Non-Line-of-Sight Identification Sinan Gezici, Hisashi Kobayashi and H. Vincent Poor Departent of Electrical Engineering School of Engineering and Applied Science Princeton University, Princeton,

More information

E0 370 Statistical Learning Theory Lecture 6 (Aug 30, 2011) Margin Analysis

E0 370 Statistical Learning Theory Lecture 6 (Aug 30, 2011) Margin Analysis E0 370 tatistical Learning Theory Lecture 6 (Aug 30, 20) Margin Analysis Lecturer: hivani Agarwal cribe: Narasihan R Introduction In the last few lectures we have seen how to obtain high confidence bounds

More information

Support Vector Machines. Machine Learning Series Jerry Jeychandra Blohm Lab

Support Vector Machines. Machine Learning Series Jerry Jeychandra Blohm Lab Support Vector Machines Machine Learning Series Jerry Jeychandra Bloh Lab Outline Main goal: To understand how support vector achines (SVMs) perfor optial classification for labelled data sets, also a

More information

Notes on Neural Networks

Notes on Neural Networks Artificial neurons otes on eural etwors Paulo Eduardo Rauber 205 Consider te data set D {(x i y i ) i { n} x i R m y i R d } Te tas of supervised learning consists on finding a function f : R m R d tat

More information

Pattern Recognition and Machine Learning. Artificial Neural networks

Pattern Recognition and Machine Learning. Artificial Neural networks Pattern Recognition and Machine Learning Jaes L. Crowley ENSIMAG 3 - MMIS Fall Seester 2016/2017 Lessons 9 11 Jan 2017 Outline Artificial Neural networks Notation...2 Convolutional Neural Networks...3

More information

5.1 The derivative or the gradient of a curve. Definition and finding the gradient from first principles

5.1 The derivative or the gradient of a curve. Definition and finding the gradient from first principles Capter 5: Dierentiation In tis capter, we will study: 51 e derivative or te gradient o a curve Deinition and inding te gradient ro irst principles 5 Forulas or derivatives 5 e equation o te tangent line

More information

Experimental Design For Model Discrimination And Precise Parameter Estimation In WDS Analysis

Experimental Design For Model Discrimination And Precise Parameter Estimation In WDS Analysis City University of New York (CUNY) CUNY Acadeic Works International Conference on Hydroinforatics 8-1-2014 Experiental Design For Model Discriination And Precise Paraeter Estiation In WDS Analysis Giovanna

More information

APPLICATION OF NEURAL NETWORK TO SHAPE OPTIMIZATION PROBLEM

APPLICATION OF NEURAL NETWORK TO SHAPE OPTIMIZATION PROBLEM Aliguliyev R.M., Majidzade K., Gasei Habasi Y. Institute of Inforation Tecnology AAS, Baku, Azerbaan Institute of Applied Mateatics Baku State University a.raiz@science.az; kabiz83@yaoo.co APPLICATIO OF

More information

Ştefan ŞTEFĂNESCU * is the minimum global value for the function h (x)

Ştefan ŞTEFĂNESCU * is the minimum global value for the function h (x) 7Applying Nelder Mead s Optiization Algorith APPLYING NELDER MEAD S OPTIMIZATION ALGORITHM FOR MULTIPLE GLOBAL MINIMA Abstract Ştefan ŞTEFĂNESCU * The iterative deterinistic optiization ethod could not

More information

A Smoothed Boosting Algorithm Using Probabilistic Output Codes

A Smoothed Boosting Algorithm Using Probabilistic Output Codes A Soothed Boosting Algorith Using Probabilistic Output Codes Rong Jin rongjin@cse.su.edu Dept. of Coputer Science and Engineering, Michigan State University, MI 48824, USA Jian Zhang jian.zhang@cs.cu.edu

More information

Estimating Parameters for a Gaussian pdf

Estimating Parameters for a Gaussian pdf Pattern Recognition and achine Learning Jaes L. Crowley ENSIAG 3 IS First Seester 00/0 Lesson 5 7 Noveber 00 Contents Estiating Paraeters for a Gaussian pdf Notation... The Pattern Recognition Proble...3

More information

The Algorithms Optimization of Artificial Neural Network Based on Particle Swarm

The Algorithms Optimization of Artificial Neural Network Based on Particle Swarm Send Orders for Reprints to reprints@benthascience.ae The Open Cybernetics & Systeics Journal, 04, 8, 59-54 59 Open Access The Algoriths Optiization of Artificial Neural Network Based on Particle Swar

More information

Ensemble Based on Data Envelopment Analysis

Ensemble Based on Data Envelopment Analysis Enseble Based on Data Envelopent Analysis So Young Sohn & Hong Choi Departent of Coputer Science & Industrial Systes Engineering, Yonsei University, Seoul, Korea Tel) 82-2-223-404, Fax) 82-2- 364-7807

More information

Machine Learning Basics: Estimators, Bias and Variance

Machine Learning Basics: Estimators, Bias and Variance Machine Learning Basics: Estiators, Bias and Variance Sargur N. srihari@cedar.buffalo.edu This is part of lecture slides on Deep Learning: http://www.cedar.buffalo.edu/~srihari/cse676 1 Topics in Basics

More information

A KERNEL APPROACH TO ESTIMATING THE DENSITY OF A CONDITIONAL EXPECTATION. Samuel G. Steckley Shane G. Henderson

A KERNEL APPROACH TO ESTIMATING THE DENSITY OF A CONDITIONAL EXPECTATION. Samuel G. Steckley Shane G. Henderson Proceedings of te 3 Winter Siulation Conference S Cick P J Sáncez D Ferrin and D J Morrice eds A KERNEL APPROACH TO ESTIMATING THE DENSITY OF A CONDITIONAL EXPECTATION Sauel G Steckley Sane G Henderson

More information

Support Vector Machines. Maximizing the Margin

Support Vector Machines. Maximizing the Margin Support Vector Machines Support vector achines (SVMs) learn a hypothesis: h(x) = b + Σ i= y i α i k(x, x i ) (x, y ),..., (x, y ) are the training exs., y i {, } b is the bias weight. α,..., α are the

More information

Feature Extraction Techniques

Feature Extraction Techniques Feature Extraction Techniques Unsupervised Learning II Feature Extraction Unsupervised ethods can also be used to find features which can be useful for categorization. There are unsupervised ethods that

More information

DEPARTMENT OF ECONOMETRICS AND BUSINESS STATISTICS

DEPARTMENT OF ECONOMETRICS AND BUSINESS STATISTICS ISSN 1440-771X AUSTRALIA DEPARTMENT OF ECONOMETRICS AND BUSINESS STATISTICS An Iproved Method for Bandwidth Selection When Estiating ROC Curves Peter G Hall and Rob J Hyndan Working Paper 11/00 An iproved

More information

A Simplified Analytical Approach for Efficiency Evaluation of the Weaving Machines with Automatic Filling Repair

A Simplified Analytical Approach for Efficiency Evaluation of the Weaving Machines with Automatic Filling Repair Proceedings of the 6th SEAS International Conference on Siulation, Modelling and Optiization, Lisbon, Portugal, Septeber -4, 006 0 A Siplified Analytical Approach for Efficiency Evaluation of the eaving

More information

NUCLEAR THERMAL-HYDRAULIC FUNDAMENTALS

NUCLEAR THERMAL-HYDRAULIC FUNDAMENTALS NUCLEAR THERMAL-HYDRAULIC FUNDAMENTALS Dr. J. Micael Doster Departent of Nuclear Engineering Nort Carolina State University Raleig, NC Copyrigted POER CYCLES Te analysis of Terodynaic Cycles is based alost

More information

A MESHSIZE BOOSTING ALGORITHM IN KERNEL DENSITY ESTIMATION

A MESHSIZE BOOSTING ALGORITHM IN KERNEL DENSITY ESTIMATION A eshsize boosting algorith in kernel density estiation A MESHSIZE BOOSTING ALGORITHM IN KERNEL DENSITY ESTIMATION C.C. Ishiekwene, S.M. Ogbonwan and J.E. Osewenkhae Departent of Matheatics, University

More information

Modular Neural Network Task Decomposition Via Entropic Clustering

Modular Neural Network Task Decomposition Via Entropic Clustering Modular Neural Network Task Decomposition Via Entropic Clustering Jorge M. Santos Instituto Superior de Engenharia do Porto Instituto de Engenharia Biomédica Porto, Portugal jms@isep.ipp.pt Joaquim Marques

More information

Pattern Classification using Simplified Neural Networks with Pruning Algorithm

Pattern Classification using Simplified Neural Networks with Pruning Algorithm Pattern Classification using Siplified Neural Networks with Pruning Algorith S. M. Karuzzaan 1 Ahed Ryadh Hasan 2 Abstract: In recent years, any neural network odels have been proposed for pattern classification,

More information

COS 424: Interacting with Data. Written Exercises

COS 424: Interacting with Data. Written Exercises COS 424: Interacting with Data Hoework #4 Spring 2007 Regression Due: Wednesday, April 18 Written Exercises See the course website for iportant inforation about collaboration and late policies, as well

More information

Multilayer Neural Networks

Multilayer Neural Networks Multilayer Neural Networs Brain s. Coputer Designed to sole logic and arithetic probles Can sole a gazillion arithetic and logic probles in an hour absolute precision Usually one ery fast procesor high

More information

An improved self-adaptive harmony search algorithm for joint replenishment problems

An improved self-adaptive harmony search algorithm for joint replenishment problems An iproved self-adaptive harony search algorith for joint replenishent probles Lin Wang School of Manageent, Huazhong University of Science & Technology zhoulearner@gail.co Xiaojian Zhou School of Manageent,

More information

Model Fitting. CURM Background Material, Fall 2014 Dr. Doreen De Leon

Model Fitting. CURM Background Material, Fall 2014 Dr. Doreen De Leon Model Fitting CURM Background Material, Fall 014 Dr. Doreen De Leon 1 Introduction Given a set of data points, we often want to fit a selected odel or type to the data (e.g., we suspect an exponential

More information

Keywords: Estimator, Bias, Mean-squared error, normality, generalized Pareto distribution

Keywords: Estimator, Bias, Mean-squared error, normality, generalized Pareto distribution Testing approxiate norality of an estiator using the estiated MSE and bias with an application to the shape paraeter of the generalized Pareto distribution J. Martin van Zyl Abstract In this work the norality

More information

Support Vector Machines. Goals for the lecture

Support Vector Machines. Goals for the lecture Support Vector Machines Mark Craven and David Page Coputer Sciences 760 Spring 2018 www.biostat.wisc.edu/~craven/cs760/ Soe of the slides in these lectures have been adapted/borrowed fro aterials developed

More information

Bloom Features. Kwabena Boahen Bioengineering Department Stanford University Stanford CA, USA

Bloom Features. Kwabena Boahen Bioengineering Department Stanford University Stanford CA, USA 2015 International Conference on Coputational Science and Coputational Intelligence Bloo Features Asok Cutkosky Coputer Science Departent Stanford University Stanford CA, USA asokc@stanford.edu Kwabena

More information

Probability Distributions

Probability Distributions Probability Distributions In Chapter, we ephasized the central role played by probability theory in the solution of pattern recognition probles. We turn now to an exploration of soe particular exaples

More information

Lecture 12: Ensemble Methods. Introduction. Weighted Majority. Mixture of Experts/Committee. Σ k α k =1. Isabelle Guyon

Lecture 12: Ensemble Methods. Introduction. Weighted Majority. Mixture of Experts/Committee. Σ k α k =1. Isabelle Guyon Lecture 2: Enseble Methods Isabelle Guyon guyoni@inf.ethz.ch Introduction Book Chapter 7 Weighted Majority Mixture of Experts/Coittee Assue K experts f, f 2, f K (base learners) x f (x) Each expert akes

More information

Introduction to Machine Learning. Recitation 11

Introduction to Machine Learning. Recitation 11 Introduction to Machine Learning Lecturer: Regev Schweiger Recitation Fall Seester Scribe: Regev Schweiger. Kernel Ridge Regression We now take on the task of kernel-izing ridge regression. Let x,...,

More information

Using EM To Estimate A Probablity Density With A Mixture Of Gaussians

Using EM To Estimate A Probablity Density With A Mixture Of Gaussians Using EM To Estiate A Probablity Density With A Mixture Of Gaussians Aaron A. D Souza adsouza@usc.edu Introduction The proble we are trying to address in this note is siple. Given a set of data points

More information

Exploring the Exponential Integrators with Krylov Subspace Algorithms for Nonlinear Circuit Simulation

Exploring the Exponential Integrators with Krylov Subspace Algorithms for Nonlinear Circuit Simulation Exploring te Exponential Integrators wit Krylov Subspace Algorits for Nonlinear Circuit Siulation Xinyuan Wang, Hao Zuang +, Cung-Kuan Ceng CSE and ECE Departents, UC San Diego, La Jolla, CA, USA + ANSYS,

More information

Bayes Decision Rule and Naïve Bayes Classifier

Bayes Decision Rule and Naïve Bayes Classifier Bayes Decision Rule and Naïve Bayes Classifier Le Song Machine Learning I CSE 6740, Fall 2013 Gaussian Mixture odel A density odel p(x) ay be ulti-odal: odel it as a ixture of uni-odal distributions (e.g.

More information

A Possible Solution to the Cosmological Constant Problem By Discrete Space-time Hypothesis

A Possible Solution to the Cosmological Constant Problem By Discrete Space-time Hypothesis A Possible Solution to te Cosological Constant Proble By Discrete Space-tie Hypotesis H.M.Mok Radiation Healt Unit, 3/F., Saiwano Healt Centre, Hong Kong SAR Got, 28 Tai Hong St., Saiwano, Hong Kong, Cina.

More information

Qualitative Modelling of Time Series Using Self-Organizing Maps: Application to Animal Science

Qualitative Modelling of Time Series Using Self-Organizing Maps: Application to Animal Science Proceedings of the 6th WSEAS International Conference on Applied Coputer Science, Tenerife, Canary Islands, Spain, Deceber 16-18, 2006 183 Qualitative Modelling of Tie Series Using Self-Organizing Maps:

More information

Use of PSO in Parameter Estimation of Robot Dynamics; Part One: No Need for Parameterization

Use of PSO in Parameter Estimation of Robot Dynamics; Part One: No Need for Parameterization Use of PSO in Paraeter Estiation of Robot Dynaics; Part One: No Need for Paraeterization Hossein Jahandideh, Mehrzad Navar Abstract Offline procedures for estiating paraeters of robot dynaics are practically

More information

1 Bounding the Margin

1 Bounding the Margin COS 511: Theoretical Machine Learning Lecturer: Rob Schapire Lecture #12 Scribe: Jian Min Si March 14, 2013 1 Bounding the Margin We are continuing the proof of a bound on the generalization error of AdaBoost

More information

Stochastic Subgradient Methods

Stochastic Subgradient Methods Stochastic Subgradient Methods Lingjie Weng Yutian Chen Bren School of Inforation and Coputer Science University of California, Irvine {wengl, yutianc}@ics.uci.edu Abstract Stochastic subgradient ethods

More information

Variational Adaptive-Newton Method

Variational Adaptive-Newton Method Variational Adaptive-Newton Method Mohaad Etiyaz Khan Wu Lin Voot Tangkaratt Zuozhu Liu Didrik Nielsen AIP, RIKEN, Tokyo Abstract We present a black-box learning ethod called the Variational Adaptive-Newton

More information

Boosting with log-loss

Boosting with log-loss Boosting with log-loss Marco Cusuano-Towner Septeber 2, 202 The proble Suppose we have data exaples {x i, y i ) i =... } for a two-class proble with y i {, }. Let F x) be the predictor function with the

More information

Polynomial Interpolation

Polynomial Interpolation Capter 4 Polynomial Interpolation In tis capter, we consider te important problem of approximatinga function fx, wose values at a set of distinct points x, x, x,, x n are known, by a polynomial P x suc

More information

OBJECTIVES INTRODUCTION

OBJECTIVES INTRODUCTION M7 Chapter 3 Section 1 OBJECTIVES Suarize data using easures of central tendency, such as the ean, edian, ode, and idrange. Describe data using the easures of variation, such as the range, variance, and

More information

A method to determine relative stroke detection efficiencies from multiplicity distributions

A method to determine relative stroke detection efficiencies from multiplicity distributions A ethod to deterine relative stroke detection eiciencies ro ultiplicity distributions Schulz W. and Cuins K. 2. Austrian Lightning Detection and Inoration Syste (ALDIS), Kahlenberger Str.2A, 90 Vienna,

More information

A KERNEL APPROACH TO ESTIMATING THE DENSITY OF A CONDITIONAL EXPECTATION. Samuel G. Steckley Shane G. Henderson

A KERNEL APPROACH TO ESTIMATING THE DENSITY OF A CONDITIONAL EXPECTATION. Samuel G. Steckley Shane G. Henderson Proceedings of te 3 Winter Siulation Conference S Cick P J Sáncez D Ferrin and D J Morrice eds A KERNEL APPROACH TO ESTIMATING THE DENSITY OF A CONDITIONAL EXPECTATION Sauel G Steckley Sane G Henderson

More information

PULSE-TRAIN BASED TIME-DELAY ESTIMATION IMPROVES RESILIENCY TO NOISE

PULSE-TRAIN BASED TIME-DELAY ESTIMATION IMPROVES RESILIENCY TO NOISE PULSE-TRAIN BASED TIME-DELAY ESTIMATION IMPROVES RESILIENCY TO NOISE 1 Nicola Neretti, 1 Nathan Intrator and 1,2 Leon N Cooper 1 Institute for Brain and Neural Systes, Brown University, Providence RI 02912.

More information

Chapter 3. Problem Solutions

Chapter 3. Problem Solutions Capter. Proble Solutions. A poton and a partile ave te sae wavelengt. Can anyting be said about ow teir linear oenta opare? About ow te poton's energy opares wit te partile's total energy? About ow te

More information

Foundations of Machine Learning Boosting. Mehryar Mohri Courant Institute and Google Research

Foundations of Machine Learning Boosting. Mehryar Mohri Courant Institute and Google Research Foundations of Machine Learning Boosting Mehryar Mohri Courant Institute and Google Research ohri@cis.nyu.edu Weak Learning Definition: concept class C is weakly PAC-learnable if there exists a (weak)

More information

A Simple Regression Problem

A Simple Regression Problem A Siple Regression Proble R. M. Castro March 23, 2 In this brief note a siple regression proble will be introduced, illustrating clearly the bias-variance tradeoff. Let Y i f(x i ) + W i, i,..., n, where

More information

Support Vector Machine Classification of Uncertain and Imbalanced data using Robust Optimization

Support Vector Machine Classification of Uncertain and Imbalanced data using Robust Optimization Recent Researches in Coputer Science Support Vector Machine Classification of Uncertain and Ibalanced data using Robust Optiization RAGHAV PAT, THEODORE B. TRAFALIS, KASH BARKER School of Industrial Engineering

More information

The Priestley-Chao Estimator

The Priestley-Chao Estimator Te Priestley-Cao Estimator In tis section we will consider te Pristley-Cao estimator of te unknown regression function. It is assumed tat we ave a sample of observations (Y i, x i ), i = 1,..., n wic are

More information

A = h w (1) Error Analysis Physics 141

A = h w (1) Error Analysis Physics 141 Introduction In all brances of pysical science and engineering one deals constantly wit numbers wic results more or less directly from experimental observations. Experimental observations always ave inaccuracies.

More information

Stationary Gaussian Markov processes as limits of stationary autoregressive time series

Stationary Gaussian Markov processes as limits of stationary autoregressive time series Stationary Gaussian Markov processes as liits of stationary autoregressive tie series Pilip A. rnst 1,, Lawrence D. Brown 2,, Larry Sepp 3,, Robert L. Wolpert 4, Abstract We consider te class, C p, of

More information

Figure 1: Equivalent electric (RC) circuit of a neurons membrane

Figure 1: Equivalent electric (RC) circuit of a neurons membrane Exercise: Leaky integrate and fire odel of neural spike generation This exercise investigates a siplified odel of how neurons spike in response to current inputs, one of the ost fundaental properties of

More information

Extension of CSRSM for the Parametric Study of the Face Stability of Pressurized Tunnels

Extension of CSRSM for the Parametric Study of the Face Stability of Pressurized Tunnels Extension of CSRSM for the Paraetric Study of the Face Stability of Pressurized Tunnels Guilhe Mollon 1, Daniel Dias 2, and Abdul-Haid Soubra 3, M.ASCE 1 LGCIE, INSA Lyon, Université de Lyon, Doaine scientifique

More information

Course Notes for EE227C (Spring 2018): Convex Optimization and Approximation

Course Notes for EE227C (Spring 2018): Convex Optimization and Approximation Course Notes for EE227C (Spring 2018): Convex Optiization and Approxiation Instructor: Moritz Hardt Eail: hardt+ee227c@berkeley.edu Graduate Instructor: Max Sichowitz Eail: sichow+ee227c@berkeley.edu October

More information

Homotopy analysis of 1D unsteady, nonlinear groundwater flow through porous media

Homotopy analysis of 1D unsteady, nonlinear groundwater flow through porous media Hootopy analysis of D unsteady, nonlinear groundwater flow troug porous edia Autor Song, Hao, Tao, Longbin Publised 7 Journal Title Journal of Coastal Researc Copyrigt Stateent 7 CERF. Te attaced file

More information

Optimal nonlinear Bayesian experimental design: an application to amplitude versus offset experiments

Optimal nonlinear Bayesian experimental design: an application to amplitude versus offset experiments Geophys. J. Int. (23) 155, 411 421 Optial nonlinear Bayesian experiental design: an application to aplitude versus offset experients Jojanneke van den Berg, 1, Andrew Curtis 2,3 and Jeannot Trapert 1 1

More information

PAC-Bayes Analysis Of Maximum Entropy Learning

PAC-Bayes Analysis Of Maximum Entropy Learning PAC-Bayes Analysis Of Maxiu Entropy Learning John Shawe-Taylor and David R. Hardoon Centre for Coputational Statistics and Machine Learning Departent of Coputer Science University College London, UK, WC1E

More information

Interactive Markov Models of Evolutionary Algorithms

Interactive Markov Models of Evolutionary Algorithms Cleveland State University EngagedScholarship@CSU Electrical Engineering & Coputer Science Faculty Publications Electrical Engineering & Coputer Science Departent 2015 Interactive Markov Models of Evolutionary

More information

Feedforward Networks

Feedforward Networks Feedforward Networks Gradient Descent Learning and Backpropagation Christian Jacob CPSC 433 Christian Jacob Dept.of Coputer Science,University of Calgary CPSC 433 - Feedforward Networks 2 Adaptive "Prograing"

More information

Research in Area of Longevity of Sylphon Scraies

Research in Area of Longevity of Sylphon Scraies IOP Conference Series: Earth and Environental Science PAPER OPEN ACCESS Research in Area of Longevity of Sylphon Scraies To cite this article: Natalia Y Golovina and Svetlana Y Krivosheeva 2018 IOP Conf.

More information

The Measurement and Evaluation of Distribution Transformer Losses Under Non-Linear Loading

The Measurement and Evaluation of Distribution Transformer Losses Under Non-Linear Loading IEEE ower Engineering Society General Meeting, Denver CO, June 9, 4 / ESGM 4-7 e Measureent and Evaluation of Distribution ransforer Losses Under Non-Linear Loading Aleksandar Danjanovic,.D., Meber IEEE

More information

3.3 Variational Characterization of Singular Values

3.3 Variational Characterization of Singular Values 3.3. Variational Characterization of Singular Values 61 3.3 Variational Characterization of Singular Values Since the singular values are square roots of the eigenvalues of the Heritian atrices A A and

More information

Warning System of Dangerous Chemical Gas in Factory Based on Wireless Sensor Network

Warning System of Dangerous Chemical Gas in Factory Based on Wireless Sensor Network 565 A publication of CHEMICAL ENGINEERING TRANSACTIONS VOL. 59, 07 Guest Editors: Zhuo Yang, Junie Ba, Jing Pan Copyright 07, AIDIC Servizi S.r.l. ISBN 978-88-95608-49-5; ISSN 83-96 The Italian Association

More information

Randomized Recovery for Boolean Compressed Sensing

Randomized Recovery for Boolean Compressed Sensing Randoized Recovery for Boolean Copressed Sensing Mitra Fatei and Martin Vetterli Laboratory of Audiovisual Counication École Polytechnique Fédéral de Lausanne (EPFL) Eail: {itra.fatei, artin.vetterli}@epfl.ch

More information

PAC-Bayesian Learning of Linear Classifiers

PAC-Bayesian Learning of Linear Classifiers Pascal Gerain Pascal.Gerain.@ulaval.ca Alexandre Lacasse Alexandre.Lacasse@ift.ulaval.ca François Laviolette Francois.Laviolette@ift.ulaval.ca Mario Marchand Mario.Marchand@ift.ulaval.ca Départeent d inforatique

More information

Least Squares Fitting of Data

Least Squares Fitting of Data Least Squares Fitting of Data David Eberly, Geoetric Tools, Redond WA 98052 https://www.geoetrictools.co/ This work is licensed under the Creative Coons Attribution 4.0 International License. To view a

More information

Inspection; structural health monitoring; reliability; Bayesian analysis; updating; decision analysis; value of information

Inspection; structural health monitoring; reliability; Bayesian analysis; updating; decision analysis; value of information Cite as: Straub D. (2014). Value of inforation analysis with structural reliability ethods. Structural Safety, 49: 75-86. Value of Inforation Analysis with Structural Reliability Methods Daniel Straub

More information

Principal Components Analysis

Principal Components Analysis Principal Coponents Analysis Cheng Li, Bingyu Wang Noveber 3, 204 What s PCA Principal coponent analysis (PCA) is a statistical procedure that uses an orthogonal transforation to convert a set of observations

More information

Deep Belief Network Training Improvement Using Elite Samples Minimizing Free Energy

Deep Belief Network Training Improvement Using Elite Samples Minimizing Free Energy Deep Belief Network Training Improvement Using Elite Samples Minimizing Free Energy Moammad Ali Keyvanrad a, Moammad Medi Homayounpour a a Laboratory for Intelligent Multimedia Processing (LIMP), Computer

More information

Numerical Solution for Non-Stationary Heat Equation in Cooling of Computer Radiator System

Numerical Solution for Non-Stationary Heat Equation in Cooling of Computer Radiator System (JZS) Journal of Zankoy Sulaiani, 9, 1(1) Part A (97-1) A119 Nuerical Solution for Non-Stationary Heat Equation in Cooling of Coputer Radiator Syste Aree A. Maad*, Faraidun K. Haa Sal**, and Najadin W.

More information

Feedforward Networks. Gradient Descent Learning and Backpropagation. Christian Jacob. CPSC 533 Winter 2004

Feedforward Networks. Gradient Descent Learning and Backpropagation. Christian Jacob. CPSC 533 Winter 2004 Feedforward Networks Gradient Descent Learning and Backpropagation Christian Jacob CPSC 533 Winter 2004 Christian Jacob Dept.of Coputer Science,University of Calgary 2 05-2-Backprop-print.nb Adaptive "Prograing"

More information

Continuity and Differentiability of the Trigonometric Functions

Continuity and Differentiability of the Trigonometric Functions [Te basis for te following work will be te definition of te trigonometric functions as ratios of te sides of a triangle inscribed in a circle; in particular, te sine of an angle will be defined to be te

More information

arxiv: v1 [cs.lg] 8 Jan 2019

arxiv: v1 [cs.lg] 8 Jan 2019 Data Masking with Privacy Guarantees Anh T. Pha Oregon State University phatheanhbka@gail.co Shalini Ghosh Sasung Research shalini.ghosh@gail.co Vinod Yegneswaran SRI international vinod@csl.sri.co arxiv:90.085v

More information

EN40: Dynamics and Vibrations. Midterm Examination Tuesday March

EN40: Dynamics and Vibrations. Midterm Examination Tuesday March EN4: Dynaics and ibrations Midter Exaination Tuesday Marc 4 14 Scool of Engineering Brown University NAME: General Instructions No collaboration of any kind is peritted on tis exaination. You ay bring

More information

Nonmonotonic Networks. a. IRST, I Povo (Trento) Italy, b. Univ. of Trento, Physics Dept., I Povo (Trento) Italy

Nonmonotonic Networks. a. IRST, I Povo (Trento) Italy, b. Univ. of Trento, Physics Dept., I Povo (Trento) Italy Storage Capacity and Dynaics of Nononotonic Networks Bruno Crespi a and Ignazio Lazzizzera b a. IRST, I-38050 Povo (Trento) Italy, b. Univ. of Trento, Physics Dept., I-38050 Povo (Trento) Italy INFN Gruppo

More information

MODIFICATION OF AN ANALYTICAL MODEL FOR CONTAINER LOADING PROBLEMS

MODIFICATION OF AN ANALYTICAL MODEL FOR CONTAINER LOADING PROBLEMS MODIFICATIO OF A AALYTICAL MODEL FOR COTAIER LOADIG PROBLEMS Reception date: DEC.99 otification to authors: 04 MAR. 2001 Cevriye GECER Departent of Industrial Engineering, University of Gazi 06570 Maltepe,

More information

Online Learning: Bandit Setting

Online Learning: Bandit Setting Online Learning: Bandit Setting Daniel asabi Summer 04 Last Update: October 0, 06 Introduction [TODO Bandits. Stocastic setting Suppose tere exists unknown distributions ν,..., ν, suc tat te loss at eac

More information