1,720,994 research outputs found
The Hidden Shape of Data: Topological Data Analysis for Anxiety Detection in Text
No full textA significant portion of the veteran population experiences post-traumatic stress disorder, or PTSD, as a result of their service. This is often accompanied by social anxiety disorder. In particular, student veterans are especially vulnerable as they struggle to integrate into a college lifestyle. In an effort to improve the support systems available for student veterans, we employ approaches from topological data analysis, an emerging area of research, to anxiety detection in text. Our models provide a tool to support psychologists and social workers in treating social anxiety.The results detailed in this paper could also have broader impacts in fields such as pedagogy and public health.Computer Scienc
Automatic Assessment of Structural Damage of Masonry Structures by Visual Analysis of Surface Cracks
No full textCrack detection on the road or building surface is normally done using manual inspection by specialists. The process consumes a lot of time, and the inspection result might differ depending on the specialist’s experience and knowledge. This work will propose an automated detection and rating of cracks on concrete surfaces based on convolutional neural networks (CNNs). Our method also provides a visualization of how the model learns the crack by directing the attention of the model to the different parts of the image by utilizing Gradient-weighted Class Activation Mapping (grad-cam) library. Finally, we show how combining two different data types, such as raw images and manually extracted features, into a hybrid convolutional neural network can increase the accuracy of the model.Computer Scienc
An Efficient Connected Components Algorithm for Massively-Parallel Devices
No full textMassively-parallel devices such as GPUs are best suited for accelerating regular algorithms. Since the memory access patterns and control flow of irregular algorithms are data dependent, such programs are more difficult to parallelize in general and a direct parallelization may not yield good performance, on GPUs in particular. However, by carefully studying the underlying problem, it may be possible to derive new algorithms that are more suitable for massively-parallel accelerators. This thesis involves studying and analyzing such an irregular algorithm, called Connected Components, and proposes an efficient algorithm, called ECL, which is faster than the existing CC algorithms on most tested inputs. Though atomic operations are fast, they can represent a bottleneck as these operations run serially and might hinder performance in the future parallel devices. This thesis also proposes a synchronous and atomic-free algorithm, called ECLaf, whose performance is comparable to the fastest existing CC algorithm.Computer Scienc
Evaluation of Cordova Accessor Host for Rapid Development of IoT Applications On Mobile Edge Devices
No full textThe Internet of Things (IoT) middleware service provides the ability for human and computers to learn and interact from billions of things that include sensors, actuators, applications, and other Internet connected objects. The realization of an edge-based IoT service framework will enable seamless integration of the Cyber-world with new physical devices and will fundamentally change and empower human interaction with the world. While there are many cloud-based IoT service frameworks, many health-care related IoT applications such as real-time fall detection systems cannot utilize cloud-based framework due to latency, privacy, and security concerns.
In this thesis, we first present an open source plug and play IoT middleware called Cordova Accessor host and we discuss the accessor design pattern, Apache Cordova and the accessor hosts for bridging the heterogeneity among IoT devices and allowing for smarter interactions, sharing and portability. We then discuss the Cordova Accessor host, an edge-based IoT middleware service framework, and a thorough analysis of the opportunities and challenges in the implementation of a fall detection application as components of accessors that embrace the heterogeneity of IoT devices and supports the composition and adaptability of IoT services. We demonstrate the reusability of accessors via building a heart rate monitoring IoT application. Finally, we shown that IoT services deployed on Android compatible devices using this framework consume around 35% less battery power than the same IoT services implemented in native Java language.Computer Scienc
Massively Parallel LZ77 Compression and Decompression on the GPU
No full textParallelizing data compression algorithms is difficult because algorithms like LZ77 have iterative dependencies that cannot easily be circumvented. Previous computation steps directly impact later steps in the algorithm, and parallelizing those dependent steps can prove difficult. My solution is to express both the encoding and decoding portions of LZ77 in terms of prefix sums, union-find operations, and other parallelizable computations. The results show that this methodology is effective in improving throughput. Compared to codes from the literature, my CUDA implementation is an order of magnitude faster but tends to have a lower compression ratio.Computer Scienc
A Massively Parallel Exact TSP Solver for Small Problem Sizes
No full textThe Traveling Salesman Problem (TSP) is a combinatorial optimization problem tasked with finding the shortest tour for visiting a set of cities such that each city is visited exactly once, and the tour ends in the starting city. This problem has gained attention among researchers because it is easy to describe yet difficult to solve. TSP has numerous important real-life applications, but its NP-hardness makes it difficult to find an optimal solution even for relatively small problem sizes. The literature describes many heuristic algorithms that solve the problem approximately but only few exact algorithms. The TSP solver implemented in this study is a GPU-accelerated exact solver for small problem sizes. The goal is to exploit the computing capabilities of modern GPUs for finding an optimal solution using simple algorithms. The algorithms used are exhaustive search for up to 7 cities and branch and bound for problems up to 30 cities. The branch and bound algorithm performs an irregular traversal of the search tree, making it challenging to parallelize efficiently, especially for massively parallel GPUs. I implemented the algorithm in CUDA and tested it on GPUs with different compute capabilities. My solver is exact and very fast. It outperforms the CONCORDE and LKH solvers for problems with up to 15 cities. On a 13-city instance, my solver is 36 and 15 times faster than CONCORDE and LKH, respectively.Computer Scienc
Enabling Transfer Learning Across Heterogeneous Domains Using Input Alignment
No full textAs High Performance Computing (HPC) architectures evolve quickly, gathering data on each architecture to create predictive performance models can be time-consuming. Instead, leveraging existing performance models for predicting the performance of a new application or an existing application on a new architecture is a better use of scientists’ time. However, such knowledge transfer in HPC can be difficult as a result of domain heterogeneity resulting from (1) data distribution shifts due to differences in architectures and (2) incomparable feature space due to differences in the names and numbers of features. Although existing transfer learning techniques can handle the first scenario (distribution shift), the second is unique to HPC and requires attention. We propose a few-shot-learning-based test-time adaptation method for Neural Network (NN)s to address this gap. The proposed approach aligns target features using a fully-connected NN to leverage existing knowledge from a pre-trained source model, and then stacking the results from the source model with a secondary model to capture the uncaptured relationships between the unique target features and labels. Our evaluations with both HPC datasets and Machine Learning (ML) benchmarks demonstrate that this approach can adapt an existing source model into a high-fidelity target model with only a few target samples. Additionally, we propose a novel distance measure to easily quantify the dissimilarity between the source and target datasets, which can help identify and explain the best model to use for knowledge transfer. Our extensive evaluations demonstrate that the proposed few-shot-learning-based test-time adaptation method augmented with input alignment achieves the lowest Mean-Square Error (MSE) when predicting a significantly different target from the source data.Computer Scienc
Evaluating CGAN Frameworks to Generate Performance Data
No full textOne of the most common challenges in implementing machine learning models, where performance is almost entirely dependent on the quantity and quality of the data available, is insufficient data. More so in the field of High-Performance Computing (HPC) where data collection is a long and a nontrivial issue. As a result, data augmentation using generative models [1] has seen a rise in the past few years. It has played a crucial role in improving the performance and accuracy of machine learning models by scaling up the data size without employing exhausting and costly processes.
A lot of the work done within this field is concentrated on image classification and segmentation. With its challenges, data augmentation has also been used in Natural Language Processing [2]. Later works on generative models gave rise data augmentation techniques like Generative Adversarial Networks (GANs) [3], followed by specialized versions like the styleGAN [2], DCGAN [3], etc. for image generation and CGAN (Conditional GAN) [4] and CTGAN [5] for regression. Our work evaluates such generative modeling frameworks in their ability of generating scientific data, specifically performance data. We also evaluate the effectiveness of these methods using open-source Machine Learning data repositories with regression datasets.Computer Scienc
Performance of Delta Huffman Compression Variants
No full textThis thesis examines an algorithm that is a combination of the Elias Delta and Huffman algorithms known as δ-Huffman. Exploiting the fact that Elias Delta codes are uniquely decodable and tend to allocate fewer bits to small integers, while Huffman coding is a lossless compression method that provides compressed symbols with a bit rate that is relatively close to the entropy of the source.
The thesis “Dynamic Unbounded Integer Compression” by Naga Sai Gowtham Mulpuri explored achievable compression ratios in the context of δ-Huffman encoding and used the frequency of inputs as the estimation for the probability distribution of the δ-Huffman algorithm on synthetic data and real-world posting list data.
Our work focuses on different assumptions about the way to estimate the probability function for the δ-Huffman algorithm. Assumptions such as using a Geometric Probability Mass Function in the estimation of the probability distribution of the δ-Huffman algorithm are made and methods such as slice and reset with the -δ Huffman algorithm are introduced.
We examine the practical and theoretical bounds on the compression capabilities of integer compression methods. Additionally, we devise methods for efficient compression of integers regardless of their specific probability distribution function. The focus of our work has been on the δ-Huffman algorithm’s compression, rather than on its computation complexity.
Twelve experiments with twelve variants of the dynamic δ-Huffman algorithm are performed. The experiments use one or more of three source data sets: synthetic data from several Geometric Probability Mass Functions (GPMF) and Poisson, real-world data from sorted inverted index gaps from Wikipedia, and a benchmark that attempts to represent realistic workload data from Silesia. The entropy estimates of these data sets and the average bit rate of the δ-Huffman algorithms offer a way to construct a comparative evolution of the performance of these algorithms.
From these experiments, the best variant out of the twelve δ-Huffman algorithms for synthetic data, real-world data, and benchmark data are determined.
Finally, given the results of the experiments, assuming GPMF without knowledge concerning the exact value of parameters, the best variant out of the twelve δ-Huffman algorithms is identified.Computer Scienc
Improving Top-N Evaluation of Recommender Systems
No full textRecommender systems are used to provide the user with a list of recommended items to help user find new items they might prefer. One of the main task of the recommender is to provide such items that the user has not seen before. But while evaluating, if the recommender correctly predicts such items we penalize the recommender, usually because the relevance of the item for that user is unknown, and because of the unknown relevance the item being recommended was not present in the test set of the recommender. In recommender systems it is very hard to get the relevance of every item for every user. In this research we are trying to address this problem by randomly adding decoys into the recommender’s test set. We will be measuring the performance of the recommender with different decoy sizes. We find that random decoys are exaggerating the advantage of popular-item recommenders, casting doubt on their usefulness.Computer Scienc
- …
