Coding for Random Projections and Approximate Near Neighbor Search

Size: px
Start display at page:

Download "Coding for Random Projections and Approximate Near Neighbor Search"

Transcription

1 Coding for Random Projetions and Approximate Near Neighbor Searh Ping Li Department of Statistis & Biostatistis Department of Computer Siene Rutgers University Pisataay, NJ 8854, USA Mihael Mitzenmaher Shool of Engineering and Applied Sienes Harvard University Cambridge, MA 238, USA Anshumali Shrivastava Department of Computer Siene Cornell University Ithaa, NY 4853, USA Abstrat This tehnial note ompares to oding (quantization) shemes for random projetions in the ontext of sub-linear time approximate near neighbor searh. The first sheme is based on uniform quantization [4] hile the seond sheme utilizes a uniform quantization plus a uniformly random offset [] (hih has been popular in pratie). The prior ork [4] ompared the to shemes in the ontext of similarity estimation and training linear lassifiers, ith the onlusion that the step of random offset is not neessary and may hurt the performane (depending on the similarity level). The task of near neighbor searh is related to similarity estimation ith importane distintions and requires on study. In this paper, e demonstrate that in the ontext of near neighbor searh, the step of random offset is not needed either and may hurt the performane (sometimes signifiantly so, depending on the similarity and other parameters). For approximate near neighbor searh, hen the target similarity level is high (e.g., orrelation >.85), our analysis suggest to use a uniform quantization to build hash tables, ith a bin idth =.5. On the other hand, hen the target similarity level is not that high, it is preferable to use larger values (e.g., 2 3). This is equivalent to say that it suffies to use only a small number of bits (or even just bit) to ode eah hashed value in the ontext of sublinear time near neighbor searh. An extensive experimental study on to reasonably large datasets onfirms the theoretial finding. Coding for building hash tables is a different task from oding for similarity estimation. For near neighbor searh, e need oding of the projeted data to determine hih bukets the data points should be plaed in (and the oded values are not stored). For similarity estimation, the purpose of oding is for aurately estimating the similarities using small storage spae. Therefore, if neessary, e an atually ode the projeted data tie (ith different bin idths). In this paper, e do not study the important issue of re-ranking of retrieved data points by using estimated similarities. That step is needed hen exat (all pairise) similarities an not be pratially stored or omputed on the fly. In a onurrent ork [5], e demonstrate that the retrieval auray an be further improved by using nonlinear estimators of the similarities based on a 2-bit oding sheme.

2 Introdution This paper fouses on the omparison of to quantization shemes for random projetions in the ontext of sublinear time near neighbor searh. The task of near neighbor searh is to identify a set of data points hih are most similar (in some measure of similarity) to a query data point. Effiient algorithms for near neighbor searh have numerous appliations in searh, databases, mahine learning, reommending systems, omputer vision, et. Developing effiient algorithms for finding near neighbors has been an ative researh topi sine the early days of modern omputing [2]. Near neighbor searh ith extremely high-dimensional data (e.g., texts or images) is still a hallenging task and an ative researh problem. Among many types of similarity measures, the (squared) Eulidian distane (denoted by d) and the orrelation (denoted by ρ) are most ommonly used. Without loss of generality, onsider to high-dimensional data vetors u, v R D. The squared Eulidean distane and orrelation are defined as follos: d = D u i v i 2, ρ = i= D i= u iv i D D i= u2 i i= v2 i In pratie, it appears that the orrelation is more often used than the distane, partly beause ρ is niely normalized ithin and. In fat, in this study, e ill assume that the marginal l 2 norms D i= u i 2 and D i= v i 2 are knon. This is a reasonable assumption. Computing the marginal l 2 norms only requires sanning the data one, hih is anyay needed during the data olletion proess. In mahine learning pratie, it is ommon to first normalize the data (to have unit l 2 norm) before feeding the data to lassifiation (e.g., SVM) or lustering (e.g., K-means) algorithms. For onveniene, throughout this paper, e assume unit l 2 norms, i.e., D i= ρ = u iv i D D D D D = u i v i, here u 2 i = vi 2 = (2) i= u2 i i= v2 i i= i= i=. Random Projetions As an effetive tool for dimensionality redution, the idea of random projetions is to multiply the data, e.g., u, v R D, ith a random normal projetion matrix R R D k (here k D), to generate: x = u R R k, y = v R R k, R = {r ij } D i= k j=, r ij N(, ) i.i.d. (3) The method of random projetions has beome popular for large-sale mahine learning appliations suh as lassifiation, regression, matrix fatorization, singular value deomposition, near neighbor searh, et. The potential benefits of oding ith a small number of bits arise beause the (unoded) projeted data, x j = D i= u ir ij and y j = D i= v ir ij, being real-valued numbers, are neither onvenient/eonomial for storage and transmission, nor ell-suited for indexing. The fous of this paper is on approximate (sublinear time) near neighbor searh in the frameork of loality sensitive hashing [3]. In partiular, e ill ompare to oding (quantization) shemes of random projetions [, 4] in the ontext of near neighbor searh..2 Uniform Quantization The reent ork [4] proposed an intuitive oding sheme, based on a simple uniform quantization: h (j) (u) = x j /, here > is the bin idth and. is the standard floor operation. () h (j) (v) = y j / (4) The folloing theorem is proved in [4] about the ollision probability P = Pr 2 ( ) h (j) (u) = h (j) (v).

3 Theorem ( ) P = Pr h (j) (u) = h (j) (v) = 2 i= (i+) i ϕ(z) In addition, P is a monotonially inreasing funtion of ρ. [ Φ ( ) ( )] (i + ) ρz i ρz Φ dz (5) ρ 2 ρ 2 The fat that P is a monotonially inreasing funtion of ρ makes (4) a suitable oding sheme for approximate near neighbor searh in the general frameork of loality sensitive hashing (LSH)..3 Uniform Quantization ith Random Offset [] proposed the folloing ell-knon oding sheme, hih uses indos and a random offset: h,q(u) (j) xj + q j =, h (j) yj + q j,q(v) = (6) here q j uniform(, ). [] shoed that the ollision probability an be ritten as ( ) ( ) ( P,q =Pr h,q(u) (j) = h,q(v) (j) t = 2ϕ t ) dt (7) d d here d = u v 2 = 2( ρ) is the Eulidean distane beteen u and v. Compared ith (6), the sheme (4) does not use the additional randomization ith q uniform(, ) (i.e., the offset). [4] elaborated the folloing advantages of (4) in the ontext of similarity estimation:. Operationally, h is simpler than h,q. 2. With a fixed, h is alays more aurate than h,q, often signifiantly so. 3. For eah oding sheme, one an separately find the optimum bin idth. The optimized h is also more aurate than optimized h,q, often signifiantly so. 4. h requires a smaller number of bits than h,q. In this paper, e ill ompare h,q ith h in the ontext of sublinear time near neighbor searh..4 Sublinear Time -Approximate Near Neighbor Searh Consider a data vetor u. Suppose there exists another vetor hose Eulidian distane ( d) from u is at most d (the target distane). The goal of -approximate d -near neighbor algorithms is to return data vetors (ith high probability) hose Eulidian distanes from u are at most d ith >. Reall that, in our definition, d = 2( ρ) is the squared Eulidian distane. To be onsistent ith [], e present the results in terms of d. Corresponding to the target distane d, the target similarity ρ an be omputed from d = 2( ρ ) i.e., ρ = d /2. To simplify the presentation, e fous on ρ (as is ommon in pratie), i.e., d 2. One e fix a target similarity ρ, an not exeed a ertain value: 2( ρ ) 2 = (8) ρ For example, hen ρ =.5, e must have 2. 3

4 Under the general frameork, the performane of an LSH algorithm largely depends on the differene (gap) beteen the to ollision probabilities P () and P (2) (respetively orresponding to d and d ): P () = Pr (h (u) = h (v)) hen d = u v 2 2 = d (9) P (2) = Pr (h (u) = h (v)) hen d = u v 2 2 = 2 d () Corresponding to h,q, the ollision probabilities P (),q and P (2),q are analogously defined. A larger differene beteen P () and P (2) implies a more effiient LSH algorithm. The folloing G values (G for h and G,q for h,q ) haraterize the gaps: G = () log /P log /P (2), G,q = log /P (),q log /P (2),q () A smaller G (i.e., larger differene beteen P () and P (2) ) leads to a potentially more effiient LSH algorithm and ρ < is partiularly desirable [3]. The general theory says the query time for -approximate d -near neighbor is dominated by O(N G ) distane evaluations, here N is the total number of data vetors in the olletion. This is better than O(N), the ost of a linear san. 2 Comparison of the Collision Probabilities To help understand the intuition hy h may lead to better performane than h,q, in this setion e examine their ollision probabilities P and P,q, hih an be expressed in terms of the standard normal pdf and df funtions: ϕ(x) = 2π e x2 2 and Φ(x) = x ϕ(x)dx, ( ) ( P,q = Pr h,q(u) (j) = h (j),q(v) = 2Φ ( ) P = Pr h (j) (u) = h (j) (v) = 2 It is lear that P,q as. i= d (i+) i ) ϕ(z) [ π/ d Φ ( (i + ) ρz ρ 2 ( / d ϕ d ) ( Φ ) i ρz ρ 2 )] (2) dz (3) Figure plots both P and P,q for seleted ρ values. The differene beteen P and P,q beomes apparent hen is not small. For example, hen ρ =, P quikly approahes the limit.5 hile P,q keeps inreasing (to ) as inreases. Intuitively, the fat that P,q hen ρ =, is undesirable beause it means to orthogonal vetors ill have the same oded value. Thus, it is not surprising that h ill have better performane than h,q, for both similarity estimation and sublinear time near neighbor searh. 4

5 Prob Prob.9 P.8.7 P,q ρ = P.6.5 P,q ρ = Prob Prob.9 P.8.7 P,q ρ = ρ =.9.8 P.7.6 P,q Prob Prob P.5 P.4,q.3.2 ρ = P.7 P,q ρ = Figure : Collision probabilities, P and P,q, for ρ =,.25,.5,.75,.9, and.99. The sheme h has smaller ollision probabilities than the sheme [] h,q, espeially hen > 2. 3 Theoretial Comparison of the s Figure 2 ompares G ith G,q at their optimum values, as funtions of, for a ide range of target similarity ρ levels. Basially, at eah and ρ, e hoose the to minimize G and the to minimize G,q. This figure illustrates that G is smaller than G,q, notieably so in the lo similarity region. Figure 3, Figure 4, Figure 5, and Figure 6 present G and G,q as funtions of, for ρ =.99, ρ =.95, ρ =.9 and ρ =.5, respetively. In eah figure, e plot the urves for a ide range of values. These figures illustrate here the optimum values are obtained. Clearly, in the high similarity region, the smallest G values are obtained at lo values, espeially at small. In the lo (or moderate) similarity region, the smallest G values are usually attained at relatively large. In pratie, e normally have to pre-speify a, for all and ρ values. In other ords, the optimum G values presented in Figure 2 are in general not attainable. Therefore, Figure 7, Figure 8, Figure 9, and Figure present G and G,q as funtions of, for ρ =.99, ρ =.95, ρ =.9 and ρ =.5, respetively. In eah figure, e plot the urves for a ide range of values. These figures again onfirm that G is smaller than G,q. 5

6 ρ =. G G,q / ρ = G G,q /.4 ρ = G G,q / G G,q /.3.2. ρ = ρ =.2 G G,q / ρ =.5 G G,q / G G,q /.4 ρ = G G,q /.3.2. ρ = G G,q /.75 ρ = ρ =.6 G G,q / G G,q /.4 ρ = G G,q /.3.2. ρ = Figure 2: Comparison of the optimum gaps (smaller the better) for h and h,q. For eah ρ and, e an find the smallest gaps individually for h and h,q, over the entire range of. We an see that for all target similarity levels ρ, both h,q and h exhibit better performane than /. h alays has smaller gap than h,q, although in high similarity region both shemes perform similarly. 6

7 ρ =.99, =.5 G G,q ρ =.99, =. G G,q.95 ρ =.99, =.2 G G,q ρ =.99, =.3 G G,q ρ =.99, =.4 G G,q ρ =.99, =.5 G G,q ρ =.99, =.7 G G,q ρ =.99, = 3 G G,q ρ =.99, = 2 G G,q G ρ.8 =.99, = 4 G,q ρ =.99, = 2.5 G G,q G ρ.8 =.99, = 5 G,q Figure 3: The gaps G and G,q as funtions of, for ρ =.99. In eah panel, e plot both G and G,q for a partiular value. The plots illustrate here the optimum values are obtained. 7

8 ρ =.95, =.5 G G,q ρ =.95, =. G G,q.95 ρ =.95, =.2 G G,q ρ =.95, =.3 G G,q ρ =.95, =.7 ρ =.95, = 3 G G,q G G,q ρ =.95, =.4 G G,q ρ =.95, = 2 G G,q ρ =.95, = 4 G.8.7 G,q ρ =.95, =.5 G G,q ρ =.95, = 2.5 G G,q ρ =.95, = 4.4 G.8.7 G,q Figure 4: The gaps G and G,q as funtions of, for ρ =.95 and a range of values. 8

9 ρ =.9, =.5 G G,q ρ =.9, =. G G,q.95 ρ =.9, =.2 G G,q ρ =.9, =.3 G G,q ρ =.9, =.7 G G,q ρ =.9, =.4 G G,q ρ =.9, = 2 G G,q ρ =.9, =.5 G G,q ρ =.9, = 2.5 G G,q Figure 5: The gaps G and G,q as funtions of, for ρ =.9 and a range of values. 9

10 ρ =.5, =.5 G G,q ρ =.5, =.3 G G,q ρ =.5, =. G G,q ρ =.5, =.35 G G,q ρ =.5, =.2 G G,q ρ =.5, =.4 G G,q Figure 6: The gaps G and G,q as funtions of, for ρ =.5 and a range of values.

11 G G,q / ρ =.99, = G G,q / ρ =.99, = G G,q / ρ =.99, = G G,q / ρ =.99, = G G,q / ρ =.99, = G G,q / ρ =.99, = G G,q / ρ =.99, = G G,q / ρ =.99, = G G,q / ρ =.99, = G G,q / ρ =.99, = G G,q / ρ =.99, = G G,q / ρ =.99, = Figure 7: The gaps G and G,q as funtions of, for ρ =.99. In eah panel, e plot both G and G,q for a partiular value.

12 G G,q / ρ =.95, = G G,q / ρ =.95, = G G,q / ρ =.95, = G G,q / ρ =.95, = G G,q / ρ =.95, = G G,q / ρ =.95, = G G,q / ρ =.95, = G G,q / ρ =.95, = G G,q / ρ =.95, = G G,q / ρ =.95, = G G,q / ρ =.95, = G G,q / ρ. =.95, = Figure 8: The gaps G and G,q as funtions of, for ρ =.95. In eah panel, e plot both G and G,q for a partiular value. 2

13 G G,q /.3 ρ =.9, = G G,q /.3 ρ =.9, = ρ =.9, = G G,q / G G,q /.3 ρ =.9, = G G,q /.3 ρ =.9, = G G,q /.3 ρ =.9, = G G,q /.3 ρ =.9, = G G,q /.3 ρ =.9, = G G,q /.3 ρ =.9, = G G,q /.3 ρ =.9, = G G,q /.3 ρ =.9, = G G,q /.3 ρ =.9, = Figure 9: The gaps G and G,q as funtions of, for ρ =.9. In eah panel, e plot both G and G,q for a partiular value. 3

14 G G,q / G G,q / G G,q /.6 ρ =.5, = ρ =.5, = G G,q / ρ =.5, =.75 ρ =.5, = 3 G G,q / G G,q / ρ =.5, = G G,q /.6 ρ =.5, = G G,q /.6 ρ =.5, = G G,q /.6 ρ =.5, = ρ =.5, = ρ =.5, =.5 G G,q / ρ =.5, = 2.5 G G,q / G G,q /.6 ρ =.5, = Figure : The gaps G and G,q as funtions of, for ρ =.5. In eah panel, e plot both G and G,q for a partiular value. 4

15 4 Optimal s To vie the optimal gaps more learly, Figure and Figure 2 plot the best gaps (left panels) and the optimal values (right panels) at hih the best gaps are attained, for seleted values of and the entire range of ρ. The results an be summarized as follos At any ρ and, the optimal gap G,q is alays at least as large as the optimal gap G. At relatively lo similarities, the optimal G,q an be substantially larger than the optimal G. When the target similarity level ρ is high (e.g., ρ >.85), for both shemes h and h,q, the optimal values are relatively lo, for example, =.5 hen.85 < ρ <.9. In this region, both h,q and h behavior similarly. When the target similarity level ρ is not so high, for h, it is best to use a large value of, in partiular 2 3. In omparison, for h,q, the optimal values gro smoothly ith dereasing ρ. These plots again onfirm the previous omparisons: (i) e should alays replae h,q ith h ; (ii) if e use h and target at very high similarity, a good hoie of might be =.5; (iii) if e use h and the target similarity is not too high, then e an safely use = 2 3. We should also mention that, although the optimal values for h appear to exhibit a jump in the right panels of Figure and Figure 2, the hoie of does not influene the performane muh, as shon in previous plots. In Figures 3 to 6, e have seen that even hen the optimal appear to approah, the atual gaps are not muh differene beteen = 3 and 3. In the real-data evaluations in the next setion, e ill see the same phenomenon for h. Note that the Gaussian density deays rapidly at the tail, for example, Φ(6) = 9.9. If e hoose =.5, or 2, or 3, then e just need a small number of bits to ode eah hashed value. 5

16 =.5 G G,q ρ =. G G,q ρ =.3 G G,q ρ Optimum Optimum Optimum =.5 2 G G,q.2.4 ρ =. 2 G G,q.2.4 ρ =.3 2 G G,q.2.4 ρ.6.8 Figure : Left panels: the optimal (smallest) gaps at given values and the entire range of ρ. We an see that G,q is alays larger than G, onfirming that it is better to use h instead of h,q. Right panels: the optimal values of at hih the optimal gaps are attained. When the target similarity ρ is very high, it is best to use a relatively small. When the target similarity is not that high, if e use h, it is best to use > 3. 6

17 = G G,q ρ =.7 G.48 G,q ρ = 2 G G,q ρ Optimum Optimum Optimum =.5 2 G G,q ρ =.7 2 G G,q ρ.9 6 = G G,q ρ.9.95 Figure 2: Left panels: the optimal (smallest) gaps at given values and the entire range of ρ. We an see that G,q is alays larger than G, onfirming that it is better to use h instead of h,q. Right panels: the optimal values of at hih the optimal gaps are attained. When the target similarity ρ is very high, it is best to use a relatively small. When the target similarity is not that high, if e use h, it is best to use > 3. 7

18 5 An Experimental Study To datasets, Peekaboom and Youtube, are used in our experiments for validating the theoretial results. Peekaboom is a standard image retrieval dataset, hih is divided into to subsets, one ith 998 data points and another ith data points. We use the larger subset for building hash tables and the smaller subset for query data points. The reported experimental results are averaged over all query data points. Available in the UCI repository, Youtube is a multi-vie dataset. For simpliity, e only use the largest set of audio features. The original training set, ith data points, is used for building hash tables. 5 data points, randomly seleted from the original test set, are used as query data points. We use the standard (K, L)-LSH implementation [3]. We generate K L independent hash funtions h i,j, i = to K, j = to L. For eah hash table j, j = to L, e onatenate K hash funtions < h,j, h 2,j, h 3,j,..., h K,j >. For eah data point, e ompute the hash values and plae them (in fat, their pointers) into the appropriate bukets of the hash table i. In the query phase, e ompute the hash value of the query data points using the same hash funtions to find the buket in hih the query data point belongs to and only searh for near neighbor among the data points in that buket of hash table i. We repeat the proess for eah hash table and the final retrieved data points are the union of the retrieved data points in all the hash tables. Ideally, the number of retrieved data points ill be substantially smaller than the total number of data points. We use the term fration retrieved to indiate the ratio of the number of retrieved data points over the total number of data points. A smaller value of fration retrieved ould be more desirable. To thoroughly evaluate the to oding shemes, e ondut extensive experiments on the to datasets, by using many ombinations of K (from 3 to 4) and L (from to 2). At eah hoie of (K, L), e vary from.5 to 5. Thus, the total number of ombinations is large, and the experiments are very time-onsuming. There are many ays to evaluate the performane of an LSH sheme. We ould speify a threshold of similarity and only ount the retrieved data points hose (exat) similarity is above the threshold as true positives. To avoid speifying a threshold and onsider the fat that in pratie people often ould like to retrieve the top-t nearest neighbors, e take a simple approah by omputing the reall based on top-t neighbors. For example, suppose the number of retrieved data points is 2, among hih 7 data points belong to the top-t. Then the reall value ould be 7/T = 7% if T =. Ideally, e hope the realls ould be as high as possible and in the meanhile e hope to keep the fration retrieved as lo as possible. Figure 3 presents the results on Youtube for T = and target realls from. to.99. In every panel, e set a target reall threshold. At every bin idth, e find the smallest fration retrieved over a ide range of LSH parameters, K and L. Note that, if the target reall is high (e.g.,.95), e basially have to effetively loer the target threshold ρ, so that e do not have to go don the re-ranked list too far. The plots sho that, for high target realls, e need to use relatively large (e.g., 2 3), and for lo target realls, e should use a relatively small (e.g., =.5). Figures 4 to 8 present similar results on the Youtube dataset for T = 5, 2,, 5, 3. We only inlude plots ith relatively high realls hih are often more useful in pratie. Figures 9 to 24 present the results on the Peekaboom dataset, hih are essentially very similar to the results on the Youtube dataset. These plots onfirm the previous theoretial analysis: (i) it is essentially alays better to use h instead of h,q, i.e., the random offset is not needed; (ii) hen using h and the target reall is high (hih essentially means hen the target similarity is lo), it is better to use a relatively large (e.g., = 2 3); (iii) hen using h and the target reall is lo, it is better to use a smaller (e.g., =.5); (iv) hen using h, the influene is is not that muh as long as it is in a reasonable range, hih is important in pratie. 8

19 Fration Retrieved Youtube: Top Reall =.99 h,q h Fration Retrieved Youtube: Top Reall =.98 h,q h Fration Retrieved Youtube: Top Reall =.97 h,q h Fration Retrieved Youtube: Top Reall =.95 h,q h Fration Retrieved Youtube: Top Reall =.93 h,q h Fration Retrieved Youtube: Top Reall =.9 h,q h Fration Retrieved Fration Retrieved Youtube: Top Reall =.85 h,q h Youtube: Top Reall =.6 h,q h Fration Retrieved Youtube: Top Reall =.8 h,q h Fration Retrieved Youtube: Top Reall =.5 h,q h Fration Retrieved Fration Retrieved Youtube: Top Reall =.7 h,q h Youtube: Top Reall =.4 h,q h Fration Retrieved Youtube: Top Reall =.3 h,q h Fration Retrieved x 3 Youtube: Top Reall = h,q h Fration Retrieved 5 x 3 Youtube: Top Reall = h,q h Figure 3: Youtube Top. In eah panel, e plot the optimal fration retrieved at a target reall value (for top-) ith respet to for both oding shemes h and h,q. Loer is better. 9

20 Fration Retrieved Youtube: Top 5 Reall =.99 h,q h Fration Retrieved Youtube: Top 5 Reall =.97 h,q h Fration Retrieved Youtube: Top 5 Reall =.95 h,q h Fration Retrieved Youtube: Top 5 Reall =.9 h,q h Fration Retrieved Fration Retrieved Youtube: Top 5 Reall =.85 h,q h Youtube: Top 5 Reall =.7 h,q h Fration Retrieved Fration Retrieved Youtube: Top 5 Reall =.8 h,q h Youtube: Top 5 Reall =.6 h,q h Figure 4: Youtube Top 5. In eah panel, e plot the optimal fration retrieved at a target reall value (for top-5) ith respet to for both oding shemes h and h,q. 2

21 Fration Retrieved Youtube: Top 2 Reall =.99 h,q h Fration Retrieved Youtube: Top 2 Reall =.97 h,q h Fration Retrieved Fration Retrieved Fration Retrieved Youtube: Top 2 Reall =.95 h,q h Youtube: Top 2 Reall =.85 Youtube: Top 2 Reall =.7 h,q h h,q h Fration Retrieved Fration Retrieved Youtube: Top 2 Reall =.9 h,q h Youtube: Top 2 Reall =.8 h,q h Fration Retrieved Youtube: Top 2 Reall =.6 h,q h Figure 5: Youtube Top 2. In eah panel, e plot the optimal fration retrieved at a target reall value (for top-2) ith respet to for both oding shemes h and h,q. 2

22 Fration Retrieved Youtube: Top Reall =.99 h,q h Fration Retrieved Youtube: Top Reall =.97 h,q h Fration Retrieved Youtube: Top Reall =.95 h,q h Fration Retrieved Youtube: Top Reall =.9 h,q h Fration Retrieved Youtube: Top Reall =.85 h,q h Fration Retrieved Youtube: Top Reall =.8 h,q h Fration Retrieved Youtube: Top Reall =.7 h,q h Fration Retrieved Youtube: Top Reall =.6 h,q h Figure 6: Youtube Top. In eah panel, e plot the optimal fration retrieved at a target reall value (for top-) ith respet to for both oding shemes h and h,q. 22

23 Fration Retrieved Youtube: Top 5 Reall =.99 h,q h Fration Retrieved Youtube: Top 5 Reall =.97 h,q h Fration Retrieved Fration Retrieved Fration Retrieved Youtube: Top 5 Reall =.95 h,q h Youtube: Top 5 Reall =.85 Youtube: Top 5 Reall =.7 h,q h h,q h Fration Retrieved Fration Retrieved Fration Retrieved Youtube: Top 5 Reall =.9 h,q h Youtube: Top 5 Reall =.8 h,q h Youtube: Top 5 Reall =.6 h,q h Figure 7: Youtube Top 5. In eah panel, e plot the optimal fration retrieved at a target reall value (for top-5) ith respet to for both oding shemes h and h,q. 23

24 Fration Retrieved Youtube: Top 3 Reall =.99 h,q h Fration Retrieved Youtube: Top 3 Reall =.97 h,q h Fration Retrieved Fration Retrieved Fration Retrieved Youtube: Top 3 Reall =.95 h,q h Youtube: Top 3 Reall =.85 Youtube: Top 3 Reall =.7 h,q h h,q h Fration Retrieved Fration Retrieved Fration Retrieved Youtube: Top 3 Reall =.9 h,q h Youtube: Top 3 Reall =.8 h,q h Youtube: Top 3 Reall =.6 h,q h Figure 8: Youtube Top 3. In eah panel, e plot the optimal fration retrieved at a target reall value (for top-3) ith respet to for both oding shemes h and h,q. 24

25 Fration Retrieved Peekaboom: Top Reall =.99 h,q h Fration Retrieved Peekaboom: Top Reall =.97 h,q h Fration Retrieved Fration Retrieved Fration Retrieved Peekaboom: Top Reall =.95 h,q h Peekaboom: Top Reall =.85 Peekaboom: Top Reall =.7 h,q h h,q h Fration Retrieved Fration Retrieved Fration Retrieved Peekaboom: Top Reall =.9 h,q h Peekaboom: Top Reall =.8 h,q h Peekaboom: Top Reall =.6 h,q h Figure 9: Peekaboom Top. In eah panel, e plot the optimal fration retrieved at a target reall value (for top-) ith respet to for both oding shemes h and h,q. 25

26 Fration Retrieved Fration Retrieved Fration Retrieved Fration Retrieved Peekaboom: Top 5 Reall =.99 h,q h Peekaboom: Top 5 Reall =.95 h,q h Peekaboom: Top 5 Reall =.85 Peekaboom: Top 5 Reall =.7 h,q h h,q h Fration Retrieved Fration Retrieved Fration Retrieved Fration Retrieved Peekaboom: Top 5 Reall =.97 h,q h Peekaboom: Top 5 Reall =.9 h,q h Peekaboom: Top 5 Reall =.8 h,q h Peekaboom: Top 5 Reall =.6 h,q h Figure 2: Peekaboom Top 5. In eah panel, e plot the optimal fration retrieved at a target reall value (for top-5) ith respet to for both oding shemes h and h,q. 26

27 Fration Retrieved.6.5 Peekaboom: Top 2 Reall =.99 h,q h Fration Retrieved Peekaboom: Top 2 Reall =.97 h,q h Fration Retrieved Fration Retrieved Fration Retrieved Peekaboom: Top 2 Reall =.95 h,q h Peekaboom: Top 2 Reall =.85 Peekaboom: Top 2 Reall =.7 h,q h h,q h Fration Retrieved Fration Retrieved Fration Retrieved Peekaboom: Top 2 Reall =.9 h,q h Peekaboom: Top 2 Reall =.8 h,q h Peekaboom: Top 2 Reall =.6 h,q h Figure 2: Peekaboom Top 2. In eah panel, e plot the optimal fration retrieved at a target reall value (for top-2) ith respet to for both oding shemes h and h,q. 27

28 Fration Retrieved Fration Retrieved Fration Retrieved Fration Retrieved Peekaboom: Top Reall =.99 h,q h Peekaboom: Top Reall =.95 h,q h Peekaboom: Top Reall =.85 Peekaboom: Top Reall =.7 h,q h h,q h Fration Retrieved Fration Retrieved Fration Retrieved Fration Retrieved Peekaboom: Top Reall =.97 h,q h Peekaboom: Top Reall =.9 h,q h Peekaboom: Top Reall =.8 h,q h Peekaboom: Top Reall =.6 h,q h Figure 22: Peekaboom Top. In eah panel, e plot the optimal fration retrieved at a target reall value (for top-) ith respet to for both oding shemes h and h,q. 28

29 Fration Retrieved Peekaboom: Top 5 Reall =.99 h,q h Fration Retrieved Peekaboom: Top 5 Reall =.97 h,q h Fration Retrieved Fration Retrieved Fration Retrieved Peekaboom: Top 5 Reall =.95 h,q h Peekaboom: Top 5 Reall =.85 Peekaboom: Top 5 Reall =.7 h,q h h,q h Fration Retrieved Fration Retrieved Fration Retrieved Peekaboom: Top 5 Reall =.9 h,q h Peekaboom: Top 5 Reall =.8 h,q h Peekaboom: Top 5 Reall =.6 h,q h Figure 23: Peekaboom Top 5. In eah panel, e plot the optimal fration retrieved at a target reall value (for top-5) ith respet to for both oding shemes h and h,q. 29

30 Fration Retrieved Peekaboom: Top 3 Reall =.99 h,q h Fration Retrieved Peekaboom: Top 3 Reall =.97 h,q h Fration Retrieved Fration Retrieved Fration Retrieved Peekaboom: Top 3 Reall =.95 h,q h Peekaboom: Top 3 Reall =.85 Peekaboom: Top 3 Reall =.7 h,q h h,q h Fration Retrieved Fration Retrieved Fration Retrieved Peekaboom: Top 3 Reall =.9 h,q h Peekaboom: Top 3 Reall =.8 h,q h Peekaboom: Top 3 Reall =.6 h,q h Figure 24: Peekaboom Top 3. In eah panel, e plot the optimal fration retrieved at a target reall value (for top-3) ith respet to for both oding shemes h and h,q. 3

31 6 Conlusion We have ompared to quantization (oding) shemes for random projetions in the ontext of sublinear time approximate near neighbor searh. The reently proposed sheme based on uniform quantization [4] is simpler than the influential existing ork [] (hih used uniform quantization ith a random offset). Our analysis onfirms that, under the general theory of LSH, the ne sheme [4] is simpler and more aurate than []. In other ords, the step of random offset in [] is not needed and may hurt the performane. Our analysis provides the pratial guidelines for using the proposed oding sheme to build hash tables. Our reommendation is to use a bin idth about =.5 hen the target similarity is high and a bin idth about = 3 hen the target similarity is not that high. In addition, using the proposed oding sheme based on uniform quantization (ithout the random offset), the influene of is not very sensitive, hih makes it very onvenient in pratial appliations. Referenes [] Mayur Datar, Niole Immorlia, Piotr Indyk, and Vahab S. Mirrokn. Loality-sensitive hashing sheme based on p-stable distributions. In SCG, pages , Brooklyn, NY, 24. [2] Jerome H. Friedman, F. Baskett, and L. Shustek. An algorithm for finding nearest neighbors. IEEE Transations on Computers, 24: 6, 975. [3] Piotr Indyk and Rajeev Motani. Approximate nearest neighbors: Toards removing the urse of dimensionality. In STOC, pages 64 63, Dallas, TX, 998. [4] Ping Li, Mihael Mitzenmaher, and Anshumali Shrivastava. Coding for random projetions. Tehnial report, arxiv:38.228, 23. [5] Ping Li, Mihael Mitzenmaher, and Anshumali Shrivastava. Coding for random projetions and nonlinear estimators. Tehnial report, 24. 3

Complexity of Regularization RBF Networks

Complexity of Regularization RBF Networks Complexity of Regularization RBF Networks Mark A Kon Department of Mathematis and Statistis Boston University Boston, MA 02215 mkon@buedu Leszek Plaskota Institute of Applied Mathematis University of Warsaw

More information

Measuring & Inducing Neural Activity Using Extracellular Fields I: Inverse systems approach

Measuring & Inducing Neural Activity Using Extracellular Fields I: Inverse systems approach Measuring & Induing Neural Ativity Using Extraellular Fields I: Inverse systems approah Keith Dillon Department of Eletrial and Computer Engineering University of California San Diego 9500 Gilman Dr. La

More information

THE EQUATION CONSIDERING CONCRETE STRENGTH AND STIRRUPS FOR DIAGONAL COMPRESSIVE CAPACITY OF RC BEAM

THE EQUATION CONSIDERING CONCRETE STRENGTH AND STIRRUPS FOR DIAGONAL COMPRESSIVE CAPACITY OF RC BEAM - Tehnial Paper - THE EQUATION CONSIDERING CONCRETE STRENGTH AND STIRRUPS FOR DIAGONAL COMPRESSIE CAPACITY OF RC BEAM Patarapol TANTIPIDOK *, Koji MATSUMOTO *, Ken WATANABE *3 and Junihiro NIWA *4 ABSTRACT

More information

Lecture 7: Sampling/Projections for Least-squares Approximation, Cont. 7 Sampling/Projections for Least-squares Approximation, Cont.

Lecture 7: Sampling/Projections for Least-squares Approximation, Cont. 7 Sampling/Projections for Least-squares Approximation, Cont. Stat60/CS94: Randomized Algorithms for Matries and Data Leture 7-09/5/013 Leture 7: Sampling/Projetions for Least-squares Approximation, Cont. Leturer: Mihael Mahoney Sribe: Mihael Mahoney Warning: these

More information

A Differential Equation for Specific Catchment Area

A Differential Equation for Specific Catchment Area Proeedings of Geomorphometry 2009. Zurih, Sitzerland, 3 ugust - 2 September, 2009 Differential Equation for Speifi Cathment rea J. C. Gallant, M. F. Huthinson 2 CSIRO Land and Water, GPO Box 666, Canberra

More information

Coding for Random Projections

Coding for Random Projections Ping Li PINGLI@STAT.RUTGERS.EDU Dept. of Statistics and Biostatistics, Dept. of omputer Science, Rutgers University, Piscataay, NJ, USA Michael Mitzenmacher MIHAELM@EES.HARVARD.EDU School of Engineering

More information

Danielle Maddix AA238 Final Project December 9, 2016

Danielle Maddix AA238 Final Project December 9, 2016 Struture and Parameter Learning in Bayesian Networks with Appliations to Prediting Breast Caner Tumor Malignany in a Lower Dimension Feature Spae Danielle Maddix AA238 Final Projet Deember 9, 2016 Abstrat

More information

CMSC 451: Lecture 9 Greedy Approximation: Set Cover Thursday, Sep 28, 2017

CMSC 451: Lecture 9 Greedy Approximation: Set Cover Thursday, Sep 28, 2017 CMSC 451: Leture 9 Greedy Approximation: Set Cover Thursday, Sep 28, 2017 Reading: Chapt 11 of KT and Set 54 of DPV Set Cover: An important lass of optimization problems involves overing a ertain domain,

More information

Millennium Relativity Acceleration Composition. The Relativistic Relationship between Acceleration and Uniform Motion

Millennium Relativity Acceleration Composition. The Relativistic Relationship between Acceleration and Uniform Motion Millennium Relativity Aeleration Composition he Relativisti Relationship between Aeleration and niform Motion Copyright 003 Joseph A. Rybzyk Abstrat he relativisti priniples developed throughout the six

More information

Assessing the Performance of a BCI: A Task-Oriented Approach

Assessing the Performance of a BCI: A Task-Oriented Approach Assessing the Performane of a BCI: A Task-Oriented Approah B. Dal Seno, L. Mainardi 2, M. Matteui Department of Eletronis and Information, IIT-Unit, Politenio di Milano, Italy 2 Department of Bioengineering,

More information

Estimating the probability law of the codelength as a function of the approximation error in image compression

Estimating the probability law of the codelength as a function of the approximation error in image compression Estimating the probability law of the odelength as a funtion of the approximation error in image ompression François Malgouyres Marh 7, 2007 Abstrat After some reolletions on ompression of images using

More information

A model for measurement of the states in a coupled-dot qubit

A model for measurement of the states in a coupled-dot qubit A model for measurement of the states in a oupled-dot qubit H B Sun and H M Wiseman Centre for Quantum Computer Tehnology Centre for Quantum Dynamis Griffith University Brisbane 4 QLD Australia E-mail:

More information

QCLAS Sensor for Purity Monitoring in Medical Gas Supply Lines

QCLAS Sensor for Purity Monitoring in Medical Gas Supply Lines DOI.56/sensoren6/P3. QLAS Sensor for Purity Monitoring in Medial Gas Supply Lines Henrik Zimmermann, Mathias Wiese, Alessandro Ragnoni neoplas ontrol GmbH, Walther-Rathenau-Str. 49a, 7489 Greifswald, Germany

More information

In Defense of MinHash Over SimHash

In Defense of MinHash Over SimHash Anshumali hrivastava Department of Computer iene Computing and Information iene Cornell University, Ithaa, NY, UA Ping Li Department of tatistis and Biostatistis Department of Computer iene Rutgers University,

More information

Computer Science 786S - Statistical Methods in Natural Language Processing and Data Analysis Page 1

Computer Science 786S - Statistical Methods in Natural Language Processing and Data Analysis Page 1 Computer Siene 786S - Statistial Methods in Natural Language Proessing and Data Analysis Page 1 Hypothesis Testing A statistial hypothesis is a statement about the nature of the distribution of a random

More information

Feature Selection by Independent Component Analysis and Mutual Information Maximization in EEG Signal Classification

Feature Selection by Independent Component Analysis and Mutual Information Maximization in EEG Signal Classification Feature Seletion by Independent Component Analysis and Mutual Information Maximization in EEG Signal Classifiation Tian Lan, Deniz Erdogmus, Andre Adami, Mihael Pavel BME Department, Oregon Health & Siene

More information

Maximum Entropy and Exponential Families

Maximum Entropy and Exponential Families Maximum Entropy and Exponential Families April 9, 209 Abstrat The goal of this note is to derive the exponential form of probability distribution from more basi onsiderations, in partiular Entropy. It

More information

Methods of evaluating tests

Methods of evaluating tests Methods of evaluating tests Let X,, 1 Xn be i.i.d. Bernoulli( p ). Then 5 j= 1 j ( 5, ) T = X Binomial p. We test 1 H : p vs. 1 1 H : p>. We saw that a LRT is 1 if t k* φ ( x ) =. otherwise (t is the observed

More information

Optimization of Statistical Decisions for Age Replacement Problems via a New Pivotal Quantity Averaging Approach

Optimization of Statistical Decisions for Age Replacement Problems via a New Pivotal Quantity Averaging Approach Amerian Journal of heoretial and Applied tatistis 6; 5(-): -8 Published online January 7, 6 (http://www.sienepublishinggroup.om/j/ajtas) doi:.648/j.ajtas.s.65.4 IN: 36-8999 (Print); IN: 36-96 (Online)

More information

A Spatiotemporal Approach to Passive Sound Source Localization

A Spatiotemporal Approach to Passive Sound Source Localization A Spatiotemporal Approah Passive Sound Soure Loalization Pasi Pertilä, Mikko Parviainen, Teemu Korhonen and Ari Visa Institute of Signal Proessing Tampere University of Tehnology, P.O.Box 553, FIN-330,

More information

The Influences of Smooth Approximation Functions for SPTSVM

The Influences of Smooth Approximation Functions for SPTSVM The Influenes of Smooth Approximation Funtions for SPTSVM Xinxin Zhang Liaoheng University Shool of Mathematis Sienes Liaoheng, 5059 P.R. China ldzhangxin008@6.om Liya Fan Liaoheng University Shool of

More information

LOGISTIC REGRESSION IN DEPRESSION CLASSIFICATION

LOGISTIC REGRESSION IN DEPRESSION CLASSIFICATION LOGISIC REGRESSIO I DEPRESSIO CLASSIFICAIO J. Kual,. V. ran, M. Bareš KSE, FJFI, CVU v Praze PCP, CS, 3LF UK v Praze Abstrat Well nown logisti regression and the other binary response models an be used

More information

c-perfect Hashing Schemes for Binary Trees, with Applications to Parallel Memories

c-perfect Hashing Schemes for Binary Trees, with Applications to Parallel Memories -Perfet Hashing Shemes for Binary Trees, with Appliations to Parallel Memories (Extended Abstrat Gennaro Cordaso 1, Alberto Negro 1, Vittorio Sarano 1, and Arnold L.Rosenberg 2 1 Dipartimento di Informatia

More information

Sensitivity Analysis in Markov Networks

Sensitivity Analysis in Markov Networks Sensitivity Analysis in Markov Networks Hei Chan and Adnan Darwihe Computer Siene Department University of California, Los Angeles Los Angeles, CA 90095 {hei,darwihe}@s.ula.edu Abstrat This paper explores

More information

Robust Recovery of Signals From a Structured Union of Subspaces

Robust Recovery of Signals From a Structured Union of Subspaces Robust Reovery of Signals From a Strutured Union of Subspaes 1 Yonina C. Eldar, Senior Member, IEEE and Moshe Mishali, Student Member, IEEE arxiv:87.4581v2 [nlin.cg] 3 Mar 29 Abstrat Traditional sampling

More information

Nonreversibility of Multiple Unicast Networks

Nonreversibility of Multiple Unicast Networks Nonreversibility of Multiple Uniast Networks Randall Dougherty and Kenneth Zeger September 27, 2005 Abstrat We prove that for any finite direted ayli network, there exists a orresponding multiple uniast

More information

Model-based mixture discriminant analysis an experimental study

Model-based mixture discriminant analysis an experimental study Model-based mixture disriminant analysis an experimental study Zohar Halbe and Mayer Aladjem Department of Eletrial and Computer Engineering, Ben-Gurion University of the Negev P.O.Box 653, Beer-Sheva,

More information

After the completion of this section the student should recall

After the completion of this section the student should recall Chapter I MTH FUNDMENTLS I. Sets, Numbers, Coordinates, Funtions ugust 30, 08 3 I. SETS, NUMERS, COORDINTES, FUNCTIONS Objetives: fter the ompletion of this setion the student should reall - the definition

More information

Average Rate Speed Scaling

Average Rate Speed Scaling Average Rate Speed Saling Nikhil Bansal David P. Bunde Ho-Leung Chan Kirk Pruhs May 2, 2008 Abstrat Speed saling is a power management tehnique that involves dynamially hanging the speed of a proessor.

More information

Bilinear Formulated Multiple Kernel Learning for Multi-class Classification Problem

Bilinear Formulated Multiple Kernel Learning for Multi-class Classification Problem Bilinear Formulated Multiple Kernel Learning for Multi-lass Classifiation Problem Takumi Kobayashi and Nobuyuki Otsu National Institute of Advaned Industrial Siene and Tehnology, -- Umezono, Tsukuba, Japan

More information

Supplementary Materials

Supplementary Materials Supplementary Materials Neural population partitioning and a onurrent brain-mahine interfae for sequential motor funtion Maryam M. Shanehi, Rollin C. Hu, Marissa Powers, Gregory W. Wornell, Emery N. Brown

More information

Likelihood-confidence intervals for quantiles in Extreme Value Distributions

Likelihood-confidence intervals for quantiles in Extreme Value Distributions Likelihood-onfidene intervals for quantiles in Extreme Value Distributions A. Bolívar, E. Díaz-Franés, J. Ortega, and E. Vilhis. Centro de Investigaión en Matemátias; A.P. 42, Guanajuato, Gto. 36; Méxio

More information

Hankel Optimal Model Order Reduction 1

Hankel Optimal Model Order Reduction 1 Massahusetts Institute of Tehnology Department of Eletrial Engineering and Computer Siene 6.245: MULTIVARIABLE CONTROL SYSTEMS by A. Megretski Hankel Optimal Model Order Redution 1 This leture overs both

More information

Lightpath routing for maximum reliability in optical mesh networks

Lightpath routing for maximum reliability in optical mesh networks Vol. 7, No. 5 / May 2008 / JOURNAL OF OPTICAL NETWORKING 449 Lightpath routing for maximum reliability in optial mesh networks Shengli Yuan, 1, * Saket Varma, 2 and Jason P. Jue 2 1 Department of Computer

More information

MAC Calculus II Summer All you need to know on partial fractions and more

MAC Calculus II Summer All you need to know on partial fractions and more MC -75-Calulus II Summer 00 ll you need to know on partial frations and more What are partial frations? following forms:.... where, α are onstants. Partial frations are frations of one of the + α, ( +

More information

Error Bounds for Context Reduction and Feature Omission

Error Bounds for Context Reduction and Feature Omission Error Bounds for Context Redution and Feature Omission Eugen Bek, Ralf Shlüter, Hermann Ney,2 Human Language Tehnology and Pattern Reognition, Computer Siene Department RWTH Aahen University, Ahornstr.

More information

SINCE Zadeh s compositional rule of fuzzy inference

SINCE Zadeh s compositional rule of fuzzy inference IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 14, NO. 6, DECEMBER 2006 709 Error Estimation of Perturbations Under CRI Guosheng Cheng Yuxi Fu Abstrat The analysis of stability robustness of fuzzy reasoning

More information

On the Licensing of Innovations under Strategic Delegation

On the Licensing of Innovations under Strategic Delegation On the Liensing of Innovations under Strategi Delegation Judy Hsu Institute of Finanial Management Nanhua University Taiwan and X. Henry Wang Department of Eonomis University of Missouri USA Abstrat This

More information

arxiv: v2 [math.pr] 9 Dec 2016

arxiv: v2 [math.pr] 9 Dec 2016 Omnithermal Perfet Simulation for Multi-server Queues Stephen B. Connor 3th Deember 206 arxiv:60.0602v2 [math.pr] 9 De 206 Abstrat A number of perfet simulation algorithms for multi-server First Come First

More information

Weighted K-Nearest Neighbor Revisited

Weighted K-Nearest Neighbor Revisited Weighted -Nearest Neighbor Revisited M. Biego University of Verona Verona, Italy Email: manuele.biego@univr.it M. Loog Delft University of Tehnology Delft, The Netherlands Email: m.loog@tudelft.nl Abstrat

More information

Modelling and Simulation. Study Support. Zora Jančíková

Modelling and Simulation. Study Support. Zora Jančíková VYSOKÁ ŠKOLA BÁŇSKÁ TECHNICKÁ UNIVERZITA OSTRAVA FAKULTA METALURGIE A MATERIÁLOVÉHO INŽENÝRSTVÍ Modelling and Simulation Study Support Zora Jančíková Ostrava 5 Title: Modelling and Simulation Code: 638-3/

More information

A Characterization of Wavelet Convergence in Sobolev Spaces

A Characterization of Wavelet Convergence in Sobolev Spaces A Charaterization of Wavelet Convergene in Sobolev Spaes Mark A. Kon 1 oston University Louise Arakelian Raphael Howard University Dediated to Prof. Robert Carroll on the oasion of his 70th birthday. Abstrat

More information

A new method of measuring similarity between two neutrosophic soft sets and its application in pattern recognition problems

A new method of measuring similarity between two neutrosophic soft sets and its application in pattern recognition problems Neutrosophi Sets and Systems, Vol. 8, 05 63 A new method of measuring similarity between two neutrosophi soft sets and its appliation in pattern reognition problems Anjan Mukherjee, Sadhan Sarkar, Department

More information

Three-dimensional morphological modelling in Delft3D-FLOW

Three-dimensional morphological modelling in Delft3D-FLOW Three-dimensional morphologial modelling in Delft3D-FLOW G. R. Lesser, J. van Kester, D.J.R. Walstra and J.A. Roelvink WL delft hydraulis email: giles.lesser@ldelft.nl Abstrat Computer modelling of sediment

More information

Connectivity and Blockage Effects in Millimeter-Wave Air-To-Everything Networks

Connectivity and Blockage Effects in Millimeter-Wave Air-To-Everything Networks 1 Connetivity and Blokage Effets in Millimeter-Wave Air-To-Everything Networks Kaifeng Han, Kaibin Huang and Robert W. Heath Jr. arxiv:1808.00144v1 [s.it] 1 Aug 2018 Abstrat Millimeter-wave (mmwave) offers

More information

The Effectiveness of the Linear Hull Effect

The Effectiveness of the Linear Hull Effect The Effetiveness of the Linear Hull Effet S. Murphy Tehnial Report RHUL MA 009 9 6 Otober 009 Department of Mathematis Royal Holloway, University of London Egham, Surrey TW0 0EX, England http://www.rhul.a.uk/mathematis/tehreports

More information

Three-dimensional Meso-scopic Analyses of Mortar and Concrete Model by Rigid Body Spring Model

Three-dimensional Meso-scopic Analyses of Mortar and Concrete Model by Rigid Body Spring Model Three-dimensional Meso-sopi Analyses of Mortar and Conrete Model by Rigid Body Spring Model K. Nagai, Y. Sato & T. Ueda Hokkaido University, Sapporo, Hokkaido, JAPAN ABSTRACT: Conrete is a heterogeneity

More information

EE 321 Project Spring 2018

EE 321 Project Spring 2018 EE 21 Projet Spring 2018 This ourse projet is intended to be an individual effort projet. The student is required to omplete the work individually, without help from anyone else. (The student may, however,

More information

FINITE WORD LENGTH EFFECTS IN DSP

FINITE WORD LENGTH EFFECTS IN DSP FINITE WORD LENGTH EFFECTS IN DSP PREPARED BY GUIDED BY Snehal Gor Dr. Srianth T. ABSTRACT We now that omputers store numbers not with infinite preision but rather in some approximation that an be paed

More information

The Laws of Acceleration

The Laws of Acceleration The Laws of Aeleration The Relationships between Time, Veloity, and Rate of Aeleration Copyright 2001 Joseph A. Rybzyk Abstrat Presented is a theory in fundamental theoretial physis that establishes the

More information

A Queueing Model for Call Blending in Call Centers

A Queueing Model for Call Blending in Call Centers A Queueing Model for Call Blending in Call Centers Sandjai Bhulai and Ger Koole Vrije Universiteit Amsterdam Faulty of Sienes De Boelelaan 1081a 1081 HV Amsterdam The Netherlands E-mail: {sbhulai, koole}@s.vu.nl

More information

A NETWORK SIMPLEX ALGORITHM FOR THE MINIMUM COST-BENEFIT NETWORK FLOW PROBLEM

A NETWORK SIMPLEX ALGORITHM FOR THE MINIMUM COST-BENEFIT NETWORK FLOW PROBLEM NETWORK SIMPLEX LGORITHM FOR THE MINIMUM COST-BENEFIT NETWORK FLOW PROBLEM Cen Çalışan, Utah Valley University, 800 W. University Parway, Orem, UT 84058, 801-863-6487, en.alisan@uvu.edu BSTRCT The minimum

More information

Performing Two-Way Analysis of Variance Under Variance Heterogeneity

Performing Two-Way Analysis of Variance Under Variance Heterogeneity Journal of Modern Applied Statistial Methods Volume Issue Artile 3 5--003 Performing Two-Way Analysis of Variane Under Variane Heterogeneity Sott J. Rihter University of North Carolina at Greensboro, sjriht@ung.edu

More information

Signals & Systems - Chapter 6

Signals & Systems - Chapter 6 Signals & Systems - Chapter 6 S. A real-valued signal x( is knon to be uniquely determined by its samples hen the sampling frequeny is s = 0,000π. For hat values of is (j) guaranteed to be zero? From the

More information

Chapter 2 Lecture 5 Longitudinal stick fixed static stability and control 2 Topics

Chapter 2 Lecture 5 Longitudinal stick fixed static stability and control 2 Topics hapter 2 eture 5 ongitudinal stik fied stati stability and ontrol 2 Topis 2.2 mg and mα as sum of the ontributions of various omponent 2.3 ontributions of ing to mg and mα 2.3.1 orretion to mα for effets

More information

Advanced Computational Fluid Dynamics AA215A Lecture 4

Advanced Computational Fluid Dynamics AA215A Lecture 4 Advaned Computational Fluid Dynamis AA5A Leture 4 Antony Jameson Winter Quarter,, Stanford, CA Abstrat Leture 4 overs analysis of the equations of gas dynamis Contents Analysis of the equations of gas

More information

Preprints of the 19th World Congress The International Federation of Automatic Control Cape Town, South Africa. August 24-29, 2014

Preprints of the 19th World Congress The International Federation of Automatic Control Cape Town, South Africa. August 24-29, 2014 Preprints of the 9th World Congress he International Federation of Automati Control Cape on, South Afria August 4-9, 4 A Step-ise sequential phase partition algorithm ith limited bathes for statistial

More information

On the Bit Error Probability of Noisy Channel Networks With Intermediate Node Encoding I. INTRODUCTION

On the Bit Error Probability of Noisy Channel Networks With Intermediate Node Encoding I. INTRODUCTION 5188 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 54, NO. 11, NOVEMBER 2008 [8] A. P. Dempster, N. M. Laird, and D. B. Rubin, Maximum likelihood estimation from inomplete data via the EM algorithm, J.

More information

Perturbation Analyses for the Cholesky Factorization with Backward Rounding Errors

Perturbation Analyses for the Cholesky Factorization with Backward Rounding Errors Perturbation Analyses for the holesky Fatorization with Bakward Rounding Errors Xiao-Wen hang Shool of omputer Siene, MGill University, Montreal, Quebe, anada, H3A A7 Abstrat. This paper gives perturbation

More information

ANALYSIS AND MODELING OF COMBINED DISCHARGE THROUGH BOX CULVERTS AND OVER CONTRACTED BROAD-CRESTED WEIRS

ANALYSIS AND MODELING OF COMBINED DISCHARGE THROUGH BOX CULVERTS AND OVER CONTRACTED BROAD-CRESTED WEIRS Seventh International Water Tehnology Conferene Egypt 1-3 April 003 ANALYSIS AND MODELING OF COMBINED DISCHARGE THROUGH BOX CULVERTS AND OVER CONTRACTED BROAD-CRESTED WEIRS Adel-Azim M. Negm Assoiate Professor,

More information

ONLINE APPENDICES for Cost-Effective Quality Assurance in Crowd Labeling

ONLINE APPENDICES for Cost-Effective Quality Assurance in Crowd Labeling ONLINE APPENDICES for Cost-Effetive Quality Assurane in Crowd Labeling Jing Wang Shool of Business and Management Hong Kong University of Siene and Tehnology Clear Water Bay Kowloon Hong Kong jwang@usthk

More information

23.1 Tuning controllers, in the large view Quoting from Section 16.7:

23.1 Tuning controllers, in the large view Quoting from Section 16.7: Lesson 23. Tuning a real ontroller - modeling, proess identifiation, fine tuning 23.0 Context We have learned to view proesses as dynami systems, taking are to identify their input, intermediate, and output

More information

Resolving RIPS Measurement Ambiguity in Maximum Likelihood Estimation

Resolving RIPS Measurement Ambiguity in Maximum Likelihood Estimation 14th International Conferene on Information Fusion Chiago, Illinois, USA, July 5-8, 011 Resolving RIPS Measurement Ambiguity in Maximum Likelihood Estimation Wenhao Li, Xuezhi Wang, and Bill Moran Shool

More information

A Functional Representation of Fuzzy Preferences

A Functional Representation of Fuzzy Preferences Theoretial Eonomis Letters, 017, 7, 13- http://wwwsirporg/journal/tel ISSN Online: 16-086 ISSN Print: 16-078 A Funtional Representation of Fuzzy Preferenes Susheng Wang Department of Eonomis, Hong Kong

More information

The Impact of Information on the Performance of an M/M/1 Queueing System

The Impact of Information on the Performance of an M/M/1 Queueing System The Impat of Information on the Performane of an M/M/1 Queueing System by Mojgan Nasiri A thesis presented to the University of Waterloo in fulfillment of the thesis requirement for the degree of Master

More information

Probabilistic and nondeterministic aspects of Anonymity 1

Probabilistic and nondeterministic aspects of Anonymity 1 MFPS XX1 Preliminary Version Probabilisti and nondeterministi aspets of Anonymity 1 Catusia Palamidessi 2 INRIA and LIX Éole Polytehnique, Rue de Salay, 91128 Palaiseau Cedex, FRANCE Abstrat Anonymity

More information

REFINED UPPER BOUNDS FOR THE LINEAR DIOPHANTINE PROBLEM OF FROBENIUS. 1. Introduction

REFINED UPPER BOUNDS FOR THE LINEAR DIOPHANTINE PROBLEM OF FROBENIUS. 1. Introduction Version of 5/2/2003 To appear in Advanes in Applied Mathematis REFINED UPPER BOUNDS FOR THE LINEAR DIOPHANTINE PROBLEM OF FROBENIUS MATTHIAS BECK AND SHELEMYAHU ZACKS Abstrat We study the Frobenius problem:

More information

Chapter 8 Hypothesis Testing

Chapter 8 Hypothesis Testing Leture 5 for BST 63: Statistial Theory II Kui Zhang, Spring Chapter 8 Hypothesis Testing Setion 8 Introdution Definition 8 A hypothesis is a statement about a population parameter Definition 8 The two

More information

DIGITAL DISTANCE RELAYING SCHEME FOR PARALLEL TRANSMISSION LINES DURING INTER-CIRCUIT FAULTS

DIGITAL DISTANCE RELAYING SCHEME FOR PARALLEL TRANSMISSION LINES DURING INTER-CIRCUIT FAULTS CHAPTER 4 DIGITAL DISTANCE RELAYING SCHEME FOR PARALLEL TRANSMISSION LINES DURING INTER-CIRCUIT FAULTS 4.1 INTRODUCTION Around the world, environmental and ost onsiousness are foring utilities to install

More information

Exploring the feasibility of on-site earthquake early warning using close-in records of the 2007 Noto Hanto earthquake

Exploring the feasibility of on-site earthquake early warning using close-in records of the 2007 Noto Hanto earthquake Exploring the feasibility of on-site earthquake early warning using lose-in reords of the 2007 Noto Hanto earthquake Yih-Min Wu 1 and Hiroo Kanamori 2 1. Department of Geosienes, National Taiwan University,

More information

Modeling of discrete/continuous optimization problems: characterization and formulation of disjunctions and their relaxations

Modeling of discrete/continuous optimization problems: characterization and formulation of disjunctions and their relaxations Computers and Chemial Engineering (00) 4/448 www.elsevier.om/loate/omphemeng Modeling of disrete/ontinuous optimization problems: haraterization and formulation of disjuntions and their relaxations Aldo

More information

The transition between quasi-static and fully dynamic for interfaces

The transition between quasi-static and fully dynamic for interfaces Physia D 198 (24) 136 147 The transition between quasi-stati and fully dynami for interfaes G. Caginalp, H. Merdan Department of Mathematis, University of Pittsburgh, Pittsburgh, PA 1526, USA Reeived 6

More information

Predicting the confirmation time of Bitcoin transactions

Predicting the confirmation time of Bitcoin transactions Prediting the onfirmation time of Bitoin transations D.T. Koops Korteweg-de Vries Institute, University of Amsterdam arxiv:189.1596v1 [s.dc] 21 Sep 218 September 28, 218 Abstrat We study the probabilisti

More information

A simple expression for radial distribution functions of pure fluids and mixtures

A simple expression for radial distribution functions of pure fluids and mixtures A simple expression for radial distribution funtions of pure fluids and mixtures Enrio Matteoli a) Istituto di Chimia Quantistia ed Energetia Moleolare, CNR, Via Risorgimento, 35, 56126 Pisa, Italy G.

More information

Planning with Uncertainty in Position: an Optimal Planner

Planning with Uncertainty in Position: an Optimal Planner Planning with Unertainty in Position: an Optimal Planner Juan Pablo Gonzalez Anthony (Tony) Stentz CMU-RI -TR-04-63 The Robotis Institute Carnegie Mellon University Pittsburgh, Pennsylvania 15213 Otober

More information

Sensor Network Localisation with Wrapped Phase Measurements

Sensor Network Localisation with Wrapped Phase Measurements Sensor Network Loalisation with Wrapped Phase Measurements Wenhao Li #1, Xuezhi Wang 2, Bill Moran 2 # Shool of Automation, Northwestern Polytehnial University, Xian, P.R.China. 1. wenhao23@mail.nwpu.edu.n

More information

Post-Mounted Delineators and Raised Pavement Markers: Their Effect on Vehicle Operations at Horizontal Curves ~n Two-Lane Rural Highways

Post-Mounted Delineators and Raised Pavement Markers: Their Effect on Vehicle Operations at Horizontal Curves ~n Two-Lane Rural Highways TRANSPORTATION RESEARCH RECORD 1324 59 Post-Mounted Delineators and Raised Pavement Markers: Their Effet on Vehile Operations at Horiontal Curves n To-Lane Rural Highays RAYMOND A. KRAMMES AND KEVIN D.

More information

Remark 4.1 Unlike Lyapunov theorems, LaSalle s theorem does not require the function V ( x ) to be positive definite.

Remark 4.1 Unlike Lyapunov theorems, LaSalle s theorem does not require the function V ( x ) to be positive definite. Leture Remark 4.1 Unlike Lyapunov theorems, LaSalle s theorem does not require the funtion V ( x ) to be positive definite. ost often, our interest will be to show that x( t) as t. For that we will need

More information

COMBINED PROBE FOR MACH NUMBER, TEMPERATURE AND INCIDENCE INDICATION

COMBINED PROBE FOR MACH NUMBER, TEMPERATURE AND INCIDENCE INDICATION 4 TH INTERNATIONAL CONGRESS OF THE AERONAUTICAL SCIENCES COMBINED PROBE FOR MACH NUMBER, TEMPERATURE AND INCIDENCE INDICATION Jiri Nozika*, Josef Adame*, Daniel Hanus** *Department of Fluid Dynamis and

More information

Design and Development of Three Stages Mixed Sampling Plans for Variable Attribute Variable Quality Characteristics

Design and Development of Three Stages Mixed Sampling Plans for Variable Attribute Variable Quality Characteristics International Journal of Statistis and Systems ISSN 0973-2675 Volume 12, Number 4 (2017), pp. 763-772 Researh India Publiations http://www.ripubliation.om Design and Development of Three Stages Mixed Sampling

More information

Determination of the reaction order

Determination of the reaction order 5/7/07 A quote of the wee (or amel of the wee): Apply yourself. Get all the eduation you an, but then... do something. Don't just stand there, mae it happen. Lee Iaoa Physial Chemistry GTM/5 reation order

More information

Optimal control of solar energy systems

Optimal control of solar energy systems Optimal ontrol of solar energy systems Viorel Badesu Candida Oanea Institute Polytehni University of Buharest Contents. Optimal operation - systems with water storage tanks 2. Sizing solar olletors 3.

More information

Variation Based Online Travel Time Prediction Using Clustered Neural Networks

Variation Based Online Travel Time Prediction Using Clustered Neural Networks Variation Based Online Travel Time Predition Using lustered Neural Networks Jie Yu, Gang-Len hang, H.W. Ho and Yue Liu Abstrat-This paper proposes a variation-based online travel time predition approah

More information

A Heuristic Approach for Design and Calculation of Pressure Distribution over Naca 4 Digit Airfoil

A Heuristic Approach for Design and Calculation of Pressure Distribution over Naca 4 Digit Airfoil IOSR Journal of Engineering (IOSRJEN) ISSN (e): 2250-3021, ISSN (p): 2278-8719 PP 11-15 www.iosrjen.org A Heuristi Approah for Design and Calulation of Pressure Distribution over Naa 4 Digit Airfoil G.

More information

CONTROL OF THERMAL CRACKING USING HEAT OF CEMENT HYDRATION IN MASSIVE CONCRETE STRUCTURES

CONTROL OF THERMAL CRACKING USING HEAT OF CEMENT HYDRATION IN MASSIVE CONCRETE STRUCTURES CONROL OF HERMAL CRACKING USING HEA OF CEMEN HYDRAION IN MASSIVE CONCREE SRUCURES. Mizobuhi (1), G. Sakai (),. Ohno () and S. Matsumoto () (1) Department of Civil and Environmental Engineering, HOSEI University,

More information

Gluing Potential Energy Surfaces with Rare Event Simulations

Gluing Potential Energy Surfaces with Rare Event Simulations This is an open aess artile published under an ACS AuthorChoie Liense, whih permits opying and redistribution of the artile or any adaptations for non-ommerial purposes. pubs.as.org/jctc Gluing Potential

More information

Evaluation of effect of blade internal modes on sensitivity of Advanced LIGO

Evaluation of effect of blade internal modes on sensitivity of Advanced LIGO Evaluation of effet of blade internal modes on sensitivity of Advaned LIGO T0074-00-R Norna A Robertson 5 th Otober 00. Introdution The urrent model used to estimate the isolation ahieved by the quadruple

More information

Phase Diffuser at the Transmitter for Lasercom Link: Effect of Partially Coherent Beam on the Bit-Error Rate.

Phase Diffuser at the Transmitter for Lasercom Link: Effect of Partially Coherent Beam on the Bit-Error Rate. Phase Diffuser at the Transmitter for Laserom Link: Effet of Partially Coherent Beam on the Bit-Error Rate. O. Korotkova* a, L. C. Andrews** a, R. L. Phillips*** b a Dept. of Mathematis, Univ. of Central

More information

Modeling Probabilistic Measurement Correlations for Problem Determination in Large-Scale Distributed Systems

Modeling Probabilistic Measurement Correlations for Problem Determination in Large-Scale Distributed Systems 009 9th IEEE International Conferene on Distributed Computing Systems Modeling Probabilisti Measurement Correlations for Problem Determination in Large-Sale Distributed Systems Jing Gao Guofei Jiang Haifeng

More information

General probability weighted moments for the three-parameter Weibull Distribution and their application in S-N curves modelling

General probability weighted moments for the three-parameter Weibull Distribution and their application in S-N curves modelling General probability weighted moments for the three-parameter Weibull Distribution and their appliation in S-N urves modelling Paul Dario Toasa Caiza a,, Thomas Ummenhofer a a KIT Stahl- und Leihtbau, Versuhsanstalt

More information

Controller Design Based on Transient Response Criteria. Chapter 12 1

Controller Design Based on Transient Response Criteria. Chapter 12 1 Controller Design Based on Transient Response Criteria Chapter 12 1 Desirable Controller Features 0. Stable 1. Quik responding 2. Adequate disturbane rejetion 3. Insensitive to model, measurement errors

More information

Development of Fuzzy Extreme Value Theory. Populations

Development of Fuzzy Extreme Value Theory. Populations Applied Mathematial Sienes, Vol. 6, 0, no. 7, 58 5834 Development of Fuzzy Extreme Value Theory Control Charts Using α -uts for Sewed Populations Rungsarit Intaramo Department of Mathematis, Faulty of

More information

arxiv:gr-qc/ v2 6 Feb 2004

arxiv:gr-qc/ v2 6 Feb 2004 Hubble Red Shift and the Anomalous Aeleration of Pioneer 0 and arxiv:gr-q/0402024v2 6 Feb 2004 Kostadin Trenčevski Faulty of Natural Sienes and Mathematis, P.O.Box 62, 000 Skopje, Maedonia Abstrat It this

More information

Relativistic Dynamics

Relativistic Dynamics Chapter 7 Relativisti Dynamis 7.1 General Priniples of Dynamis 7.2 Relativisti Ation As stated in Setion A.2, all of dynamis is derived from the priniple of least ation. Thus it is our hore to find a suitable

More information

On the Designs and Challenges of Practical Binary Dirty Paper Coding

On the Designs and Challenges of Practical Binary Dirty Paper Coding On the Designs and Challenges of Pratial Binary Dirty Paper Coding 04 / 08 / 2009 Gyu Bum Kyung and Chih-Chun Wang Center for Wireless Systems and Appliations Shool of Eletrial and Computer Eng. Outline

More information

Normative and descriptive approaches to multiattribute decision making

Normative and descriptive approaches to multiattribute decision making De. 009, Volume 8, No. (Serial No.78) China-USA Business Review, ISSN 57-54, USA Normative and desriptive approahes to multiattribute deision making Milan Terek (Department of Statistis, University of

More information

UPPER-TRUNCATED POWER LAW DISTRIBUTIONS

UPPER-TRUNCATED POWER LAW DISTRIBUTIONS Fratals, Vol. 9, No. (00) 09 World Sientifi Publishing Company UPPER-TRUNCATED POWER LAW DISTRIBUTIONS STEPHEN M. BURROUGHS and SARAH F. TEBBENS College of Marine Siene, University of South Florida, St.

More information

THEORETICAL ANALYSIS OF EMPIRICAL RELATIONSHIPS FOR PARETO- DISTRIBUTED SCIENTOMETRIC DATA Vladimir Atanassov, Ekaterina Detcheva

THEORETICAL ANALYSIS OF EMPIRICAL RELATIONSHIPS FOR PARETO- DISTRIBUTED SCIENTOMETRIC DATA Vladimir Atanassov, Ekaterina Detcheva International Journal "Information Models and Analyses" Vol.1 / 2012 271 THEORETICAL ANALYSIS OF EMPIRICAL RELATIONSHIPS FOR PARETO- DISTRIBUTED SCIENTOMETRIC DATA Vladimir Atanassov, Ekaterina Detheva

More information

arxiv: v1 [cs.ai] 4 Sep 2015

arxiv: v1 [cs.ai] 4 Sep 2015 Quantization based Fast Inner Produt Searh Ruiqi Guo, Sanjiv Kumar, Krzysztof Choromanski, David Simha Google Researh, New York, NY, USA {guorq, sanjivk, khoro, dsimha}@google.om arxiv:59.469v [s.ai] 4

More information

The Second Postulate of Euclid and the Hyperbolic Geometry

The Second Postulate of Euclid and the Hyperbolic Geometry 1 The Seond Postulate of Eulid and the Hyperboli Geometry Yuriy N. Zayko Department of Applied Informatis, Faulty of Publi Administration, Russian Presidential Aademy of National Eonomy and Publi Administration,

More information