#### DMCA

## Efficient Computation of kNearest Neighbour Graphs for (2013)

Venue: | Large High-Dimensional Data Sets on GPU Clusters |

Citations: | 1 - 0 self |

### Citations

86 | Random projection trees and low dimensional manifolds. STOC,
- Dasgupta, Freund
- 2008
(Show Context)
Citation Context ...or every object in the input data once a convenient search indexing data structure has been built. Such search data structures include kd-trees [9], BBD-trees [10], random-projection trees (rp-trees) =-=[11]-=-, and hashing based on locally sensitive hash [12]. These method focus on optimizing the k-NN search, i.e., finding k-NNs for a set of query points w.r.t. a set of points with which the search data st... |

28 | Fast approximate kNN graph construction for high dimensional data via recursive Lanczos bisection. - Chen, Fang, et al. - 2009 |

27 | Efficient k-nearest neighbor graph construction for generic similarity measures. - Dong, Charikar, et al. - 2011 |

26 | Fast construction of k-nearest neighbor graphs for point clouds
- Connor
- 2010
(Show Context)
Citation Context ...l in the dimenision d of the data. The same exponential dependence on dimension of time complexity is seen in methods based on space filling curves such as the Hilbert’s curve [17] and Morton’s curve =-=[18]-=-. There are several approximate methods that can handle moderately high dimensional data, typically with a trade-off between speed and accuracy. One set of techniques typically is based on a hybrid of... |

15 | Scalable k-NN graph construction for visual descriptors. - Wang, Wang, et al. - 2012 |

10 |
JC (2000) A global geometric framework for nonlinear dimensionality reduction
- JB, deSilva, et al.
(Show Context)
Citation Context ...ing interests exist. * E-mail: dsouza@uwm.edu Introduction K-Nearest neighbor graphs have a variety of applications in bioinformatics [1,2], data mining [3], machine learning [4,5], manifold learning =-=[6]-=-, clustering analysis [7], and pattern recognition [8]. The main reason behind the popularity of neighborhood graphs lies in their ability to extract underlying information about structure and governi... |

8 |
Automatically tuned linear algebra software. SC98: High Performance Networking and Computing
- RC, Dongarra
- 1998
(Show Context)
Citation Context ...puting language. The MATLAB implementation took 56 hours on an exclusive CPU cluster with 32 nodes for two million images of diffraction patterns with the use of highly a optimized ATLAS-BLAS library =-=[33]-=- for multi-threaded Matrix-Matrix Multiplication in double precision. The cluster had one Xeon E5420 quad-core CPU per node with 16 kB of L1 cache, 6144 kB or L2 cache and 40 GFLOPS of double precisio... |

7 | A practical gpu based knn algorithm
- Kuang, Zhao
- 2009
(Show Context)
Citation Context ... memory write as in [25]. Finally, for kw192 it may not even be possible to execute the method in share memory even on the latest GPUs. All these methods dramatically lose performance for large k. In =-=[27]-=- a radix sort based approach is used to select the k nearest neigbhbors. They claim that for large data sets, especially for a large number of queries, the selection process dominates. A simple comple... |

6 |
von Luxburg U (2009) Optimal construction of k-nearestneighbor graphs for identifying noisy clusters. Theor Comput Sci 410: 1749
- Maier, Hein
(Show Context)
Citation Context ...e declared that no competing interests exist. * E-mail: dsouza@uwm.edu Introduction K-Nearest neighbor graphs have a variety of applications in bioinformatics [1,2], data mining [3], machine learning =-=[4,5]-=-, manifold learning [6], clustering analysis [7], and pattern recognition [8]. The main reason behind the popularity of neighborhood graphs lies in their ability to extract underlying information abou... |

6 |
An O(nlogn) algorithm for the all-nearestneighbor problem.
- PM
- 1989
(Show Context)
Citation Context ...ucting an exact k-NNGs has been investigated extensively to avoid the O n2 complexity of the brute force method. An O(n logd{1 n) was presented in [13]. [14] presented a O(cd log n) algorithm and =-=[15]-=- presented a worst case O((c0d)dn log n). In [16] two practical algorithms for k-NNGs based on recursive partitioning and pivoting have empirically shown time complexities 0:685e0:23dn1:48 and 0:858e0... |

6 |
Kandrot E (2010) CUDA by Example: An Introduction to GeneralPurpose GPU Programming
- Sanders
(Show Context)
Citation Context ...osone.org 9 September 2013 | Volume 8 | Issue 9 | e74113 Each of these tasks is coded as kernels. In our clusters we have Tesla C2050 compute GPUs from NVIDIA. We use the CUDA programming environment =-=[31]-=- to code our kernels. Finding vector norms. Finding vector B̂ of input data norms is done once in the beginning at the same time that the input files AJ are communicated to each node. For example, if ... |

5 |
CS, Noble WS (2004) Protein ranking: from local to global structure in the protein similarity network. Proc Natl Acad Sci U S A 101: 6559–6563
- Weston, Elisseeff, et al.
(Show Context)
Citation Context ...ipt. Competing Interests: The authors have declared that no competing interests exist. * E-mail: dsouza@uwm.edu Introduction K-Nearest neighbor graphs have a variety of applications in bioinformatics =-=[1,2]-=-, data mining [3], machine learning [4,5], manifold learning [6], clustering analysis [7], and pattern recognition [8]. The main reason behind the popularity of neighborhood graphs lies in their abili... |

5 |
Parallel Programming with MPI
- PS
- 1996
(Show Context)
Citation Context ... : J%Q~q. For each block row I , the partition AI is read in parts in parallel by all nodes from the shared drive. All nodes then share the parts that they each read through an ‘all_gather’ operation =-=[29]-=- to build a local copy of AI in memory. Each node q then reads AJ jJ : J%Q~q from its local drive. Since the local disk read operation is slower, we use a memory buffer and overlap computation and dis... |

4 |
Fast algorithms for the all nearest neighbors problem
- KL
- 1983
(Show Context)
Citation Context ...onstruction directly. The problem of constructing an exact k-NNGs has been investigated extensively to avoid the O n2 complexity of the brute force method. An O(n logd{1 n) was presented in [13]. =-=[14]-=- presented a O(cd log n) algorithm and [15] presented a worst case O((c0d)dn log n). In [16] two practical algorithms for k-NNGs based on recursive partitioning and pivoting have empirically shown tim... |

3 |
Rusyn I, et al. (2007) Inferring missing genotypes in large snp panels using fast nearest-neighbor searches over sliding windows
- Roberts, McMillan, et al.
(Show Context)
Citation Context ...ipt. Competing Interests: The authors have declared that no competing interests exist. * E-mail: dsouza@uwm.edu Introduction K-Nearest neighbor graphs have a variety of applications in bioinformatics =-=[1,2]-=-, data mining [3], machine learning [4,5], manifold learning [6], clustering analysis [7], and pattern recognition [8]. The main reason behind the popularity of neighborhood graphs lies in their abili... |

3 |
Multidimensional divide-and-conquer
- JL
- 1980
(Show Context)
Citation Context ...-NNG construction directly. The problem of constructing an exact k-NNGs has been investigated extensively to avoid the O n2 complexity of the brute force method. An O(n logd{1 n) was presented in =-=[13]-=-. [14] presented a O(cd log n) algorithm and [15] presented a worst case O((c0d)dn log n). In [16] two practical algorithms for k-NNGs based on recursive partitioning and pivoting have empirically sho... |

3 |
Barlaud M (2010) K-nearest neighbor search: Fast GPU-based implementations and application to high-dimensional feature matching
- Garcia, Debreuve, et al.
(Show Context)
Citation Context ... several methods that accelerate brute force k-NN and k-NNGs on graphics processing units. They primarily differ in the manner of selecting k smallest elements in every row in the distance matrix. In =-=[23,24]-=- each row of the distance matrix is processed by one thread. Each thread uses an modified insertion sort algorithm to select the k nearest elements. The number of steps that each thread takes to proce... |

3 |
Moscato P (2012) GPU-FS-kNN: A software tool for fast and scalable kNN computation using GPUs. PLoS One 7: e44000
- AS, Riveros, et al.
(Show Context)
Citation Context ... selection breaks even with the work of [23] at about k~128 and outperforms it by 15| for k~1024. Overall our implementation has a performance advantage of 7:87|. Figure 8(b) shows the speedup versus =-=[24]-=-. For this test we re-formulated the Pearson distance computation to enable the use of optimized matrix-matrix multiplication. Consequently, our distance implementation has a roughly 9| performance ad... |

2 |
Chang SF (2010) Large graph construction for scalable semisupervised learning
- Liu, He
(Show Context)
Citation Context ...e declared that no competing interests exist. * E-mail: dsouza@uwm.edu Introduction K-Nearest neighbor graphs have a variety of applications in bioinformatics [1,2], data mining [3], machine learning =-=[4,5]-=-, manifold learning [6], clustering analysis [7], and pattern recognition [8]. The main reason behind the popularity of neighborhood graphs lies in their ability to extract underlying information abou... |

2 |
Virmajoki O, Hautamäki V (2006) Fast agglomerative clustering using a k-nearest neighbor graph
- Fränti
(Show Context)
Citation Context ...mail: dsouza@uwm.edu Introduction K-Nearest neighbor graphs have a variety of applications in bioinformatics [1,2], data mining [3], machine learning [4,5], manifold learning [6], clustering analysis =-=[7]-=-, and pattern recognition [8]. The main reason behind the popularity of neighborhood graphs lies in their ability to extract underlying information about structure and governing dimensions of data clo... |

2 |
Osipov A, Rokhlin V (2011) Randomized approximate nearest neighbors algorithm
- PW
(Show Context)
Citation Context ...NNG can be built by repeatedly applying the k-NN query for every object in the input data once a convenient search indexing data structure has been built. Such search data structures include kd-trees =-=[9]-=-, BBD-trees [10], random-projection trees (rp-trees) [11], and hashing based on locally sensitive hash [12]. These method focus on optimizing the k-NN search, i.e., finding k-NNs for a set of query po... |

2 |
Immorlica N, Indyk P, Mirrokni VS (2004) Locality-sensitive hashing scheme based on p-stable distributions
- Datar
(Show Context)
Citation Context ...t search indexing data structure has been built. Such search data structures include kd-trees [9], BBD-trees [10], random-projection trees (rp-trees) [11], and hashing based on locally sensitive hash =-=[12]-=-. These method focus on optimizing the k-NN search, i.e., finding k-NNs for a set of query points w.r.t. a set of points with which the search data structure is built, ignoring the fact that every que... |

2 |
Chávez E, Figueroa K, Navarro G (2006) Practical construction of knearest neighbor graphs in metric spaces
- Paredes
(Show Context)
Citation Context ...nsively to avoid the O n2 complexity of the brute force method. An O(n logd{1 n) was presented in [13]. [14] presented a O(cd log n) algorithm and [15] presented a worst case O((c0d)dn log n). In =-=[16]-=- two practical algorithms for k-NNGs based on recursive partitioning and pivoting have empirically shown time complexities 0:685e0:23dn1:48 and 0:858e0:11dn2:15, respectively. All these algorithms hav... |

2 | Tenllado C, Prieto-Matias M, Marin M (2011) knn query processing in metric spaces using gpus - RJ, JI |

2 |
Hosino T (2009) Solving k-nearest vector problem on multiple graphics processors. CoRR abs/0906.0231
- Kato
(Show Context)
Citation Context ...the buffer fills up, all threads synchronize and then push their elements on to the heap in a serialized manner. The last stage of this algorithm can be extremely expensive especially for large k. In =-=[26]-=-, a thread block is used to process a single row. Each thread in the thread block PLOS ONE | www.plosone.org 1 September 2013 | Volume 8 | Issue 9 | e74113 strides through the array storing the k smal... |

2 |
GN, Ourmazd A. 2010 Mapping the conformations of biological assemblies
- Schwander, Fung, et al.
(Show Context)
Citation Context ...r n that fit into GPU memory, this process underuses GPU resources. Our target application is Manifold Embedding to recover structure and conformations from a large data set of images with high noise =-=[28]-=-. The underlying assumption behind manifold embedding states that a cloud of high dimensional data makes a low dimensional hyper-surface in high dimensional space. The generated hyper surface, called ... |

1 |
Pas Rvd (2007) Using OpenMP: Portable Shared Memory Parallel Programming (Scientific and Engineering Computation
- Chapman, Jost
(Show Context)
Citation Context ...GPUs. The accumulated column k-NNs are then merged by one GPU per column. The final result of this operation is the local row and column k-NNs for the partition S(I ,J). We use OpenMP multi-threading =-=[30]-=-. Each node runs Mz1 CPU threads. M threads control the M GPUs while one thread is in charge of I/O from the local disk as well as the shared disk. Distribution of Tasks and Data within GPUs The tasks... |