A New Approach for Sparse Matrix Classification Based on Deep Learning Techniques

Abstract

In this paper, a new methodology to select the best storage format for sparse matrices based on deep learning techniques is introduced. We focus on the selection of the proper format for the sparse matrixvector multiplication (SpMV), which is one of the most important computational kernels in many scientific and engineering applications. Our approach considers the sparsity pattern of the matrices as an image, using the RGB channels to code several of the matrix properties. As a consequence, we generate image datasets that include enough information to successfully train a Convolutional Neural Network (CNN). Considering GPUs as target platforms, the trained CNN selects the best storage format 90.1% of the time, obtaining 99.4% of the highest SpMV performance among the tested formats.

Sparse Matrix Classification on Imbalanced Datasets Using Convolutional Neural Networks

“…The previous method allows to generate easily a binary image dataset that fits a CNN. However, we demonstrated that datasets of that type do not provide satisfactory classification results [3]. We must take into account that scaling down a sparse matrix simplifies the appearance of its sparsity pattern, causing a loss in the information provided to the CNN in the training phase.…”

“…This section summarizes the methodology to select the best performing format for a particular sparse matrix with the aim of maximizing its SpMV performance [3]. Figure 2 shows an scheme with the different stages of our approach.…”

“…This network has about 60 million free parameters. Although AlexNet is relatively simple with respect to other standard networks, we demonstrated in [3] that using our methodology it can successfully classify sparse matrices considering a balanced dataset and three storage formats (classes). In this paper we go a step further introducing SpNet, a new simplified version of AlexNet which is able to achieve better results than the original network considering a more complex classification scenario (higher number of classes and an imbalanced dataset).…”

“…Their values should be in the 1-255 range, so a normalization process is necessary. Details about that procedure are provided in [3]. There are several combinations of channels and metrics, but in this work we only consider those which achieve the best performance results: R 1 G 3 B 4 and R 0 G 1 B 4 .…”

“…In this way, the produced datasets can successfully train a CNN. This approach has been successfully applied to a more simple classification task with only three classes and a balanced dataset [3]. The main contributions of this work are the following:…”