Paweł Drozda

SVM with CUDA accelerated kernels for big sparse problems

The SVM algorithm is one of the most frequently used methods for the classification process. For many domains, where the classification problems have many features as well as numerous instances, classification is a difficult and time-consuming task. For this reason, the following paper presents the CSR-GPU-SVM algorithm which accelerates SVM training for large and sparse problems with the use of the CUDA technology. Implementation is based on the SMO (Sequential Minimal Optimization) algorithm and utilizes the CSR(Compressed Sparse Row) sparse matrix format. The proposed solution allows us to perform efficient classification of big datasets, for example rcv1 and newsgroup20, for which classification with dense representation is not possible. The performed experiments have proven the accelerations in the order of 6 - 35 training times compared to original LibSVM implementation.

Stochastic Gradient Descent with Barzilai-Borwein update step for SVM

This paper presents a new approach to solving the optimization task that arises when L2-SVM in its primal form is considered. In particular, we propose the application of a Barzilai-Borwein (BB) update step in five variants for the classic Stochastic Gradient Descent (SGD) algorithm. The evaluation is designed to check the effectiveness of the proposed methods in large scale scenarios in terms of execution time, convergence and sensitivity to the choice of initial parameters. The obtained results are compared with those obtained for well-known linear SVM algorithms and they indicate that the level of convergence of the proposed methods is very similar to that found in the other studies. Moreover, our approach shows much lower sensitivity to the choice of initial parameters, which allows for a substantial reduction of pre-processing.

Online Crowdsource System Supporting Ground Truth Datasets Creation

This paper proposes a design of a system for creating image similarity datasets which are necessary for testing the quality of supervised ranking algorithms. In particular, the main goal is to facilitate the creation of similar images rankings for given a imaginary dataset. The system was designed in a manner that involves user feedback in the process of creating the rankings. In each iteration of ranking construction, the query image and twelve candidates are presented to the user, who is intended to select the most similar one. Moreover, in order to accelerate the method convergence the approach based on simulated annealing is adapted. It initially chooses the images randomly from a dataset and in the later stages the images with rank rate above zero are chosen with certain probability.

Ranking by k-means voting algorithm for similar image retrieval

Recently, the field of CBIR has attracted a lot of attention in the literature. In this paper, the problem of visually similar image retrieval has been investigated. For this task we use the methods derived from the Bag of Visual Words approach, such as Scale Invariant Feature Transform (SIFT) for identifying image keypoints and K-means to build a visual dictionary. To create a ranking of similar images, a novel Ranking by K-means Voting algorithm is proposed. The experimental section shows that our method works well for similar image retrieval. It turned out that our results are more accurate in comparison with a classical similarity measure based on the Euclidean metric in the order of 6% - 15%.

Visual words sequence alignment for image classification

In recent years, the field of image processing has been gaining a growing interest in many scientific domains. In this paper, the attention is focused on one of the fundamental image processing problems, that is image classification. In particular, the novel approach of bridging content based image retrieval and sequence alignment domains was introduced. For this purpose, the dense version of the SIFT key point descriptor, k-means for visual dictionary construction and the Needleman-Wunsch method for sequence alignment were implemented. The performed experiments, which evaluated the classification accuracy, showed the great potential of the proposed solution indicating new directions for development of new image classification algorithms.

Different Orderings and Visual Sequence Alignment Algorithms for Image Classification

This paper presents a successful connection of different sequence alignment algorithms with Bag of Visual Words concept for image classification. In particular, sequences were created on the basis of dense SIFT descriptors, for which different types of sequence orderings were proposed. Then, the similarities between images were calculated with two different sequence alignment algorithms. Finally, the SVM algorithm was proposed as a classifier. The obtained results showed that both sequence alignment algorithms obtain very similar results and that the type of ordering affects the accuracy very slightly.

GPU Accelerated SVM with Sparse Sliced EllR-T Matrix Format

This paper presents the SECu-SVM algorithm for solving classification problems. It allows for a significant acceleration of the standard SVM implementations by transferring the most time-consuming computations from the standard CPU to the Graphics Processor Units (GPU). In addition, highly efficient Sliced EllR-T sparse matrix format was used for storing the dataset in GPU memory, which requires a very low memory footprint and is also well adapted to parallel processing. Performed experiments demonstrate an acceleration of 4–100 times over LibSVM. Moreover, in the majority of cases the SECu-SVM is less time-consuming than the best sparse GPU implementations and allows for handling significantly larger classification datasets.

Accelerating SVM with GPU: The State of the Art

This article summarizes the achievements that have been made in the field of GPU SVM acceleration. In particular, the algorithms which allow the acceleration of SVM classification performed on dense datasets are presented and the limitations of the dataset size are pointed out. Moreover, the solutions which deal with large sparse collections are demonstrated. These algorithms apply different sparse dataset formats to make possible the classification on the GPU. Finally, GPU implementations for different SVM kernel functions are provided.

GPU Solver with Chi-square Kernels for SVM Classification of Big Sparse Problems

This paper presents the ongoing research on the GPU SVM solutions for classification of big sparse datasets. In particular, after the success of implementation of RBF kernel for sparse matrix formats in previous work we decided to evaluate Chi

Visual words selection based on class separation measures

^{2}