Browsing by Author "Gaur, Yamini"

Now showing 1 - 2 of 2
  • Thumbnail Image
    Big Data Text Summarization - Hurricane Irma
    Chava, Raja Venkata Satya Phanindra; Dhar, Siddharth; Gaur, Yamini; Rambhakta, Pranavi; Shetty, Sourabh (Virginia Tech, 2018-12-13)
    With the increased rate of content generation on the Internet, there is a pressing need for making tools to automate the process of extracting meaningful data. Big data analytics deals with researching patterns or implicit correlations within a large collection of data. There are several sources to get data from, such as news websites, social media platforms (for example FaceBook and Twitter), sensors, and other IoT (Internet of Things) devices. Social media platforms like Twitter prove to be important sources of data collection since the level of activity increases significantly during major events such as hurricanes, floods, and events of global importance. For generating summaries, we first had to convert the WARC file which was given to us, into JSON format, which was more understandable. We then cleaned the text by removing boilerplate and redundant information. After that, we proceeded with removing stopwords and getting a collection of the most important words occurring in the documents. This ensured that the resulting summary would have important information from our corpus and would still be able to answer all the questions. One of the challenges that we faced at this point was to decide how to correlate words in order to get the most relevant words out of a document. We tried several techniques such as TF-IDF in order to resolve this. Correlation of different words with each other is an important factor in generating a cohesive summary because while a word may not be in the list of most commonly occurring words in the corpus, it could still be relevant and give significant information about the event. Due to the occurrence of Hurricane Irma around the same time as the occurrence of Hurricane Harvey, a large number of documents were not about Hurricane Irma. Due to this, all such documents were eliminated as they were deemed non-relevant. Classification of documents as relevant or non-relevant ensured that our deep learning summaries were not getting generated on data that was not crucial in building our final summary. Initially, we attempted to use Mahout classifiers, but the results obtained were not satisfactory. Instead, we used a much simpler world filtering approach for classification which has eliminated a significant number of documents by classifying them as non-relevant. We used the Pointer-Generator technique, which implements a Recurrent Neural Network (RNN) for building the deep learning abstractive summary. We combined data from multiple relevant documents into a single document, and thus generated multiple summaries, each corresponding to a set of documents. We wrote a Python script to perform post-processing on the generated summary to convert all the alphabetic characters after a period and space to uppercase. This was important because for lemmatization, stopword removal, and POS tagging, the whole dataset is converted to lowercase. The script also converts the first alphabetic character of all POS-tagged proper nouns to upper case. ROUGE (Recall-Oriented Understudy for Gisting Evaluation) is used to evaluate the generated summary against the golden standard summary. The abstractive summary returns good evaluation results when compared with the Golden Standard on the ROUGE_sent evaluation. The ROUGE_para and cov_entity evaluation results were not up to the mark, but we feel that was mainly due to the writing style of the Gold Standard as our abstractive summary was able provide most of the information related to Hurricane Irma.
  • Thumbnail Image
    Exploring Per-Input Filter Selection and Approximation Techniques for Deep Neural Networks
    Gaur, Yamini (Virginia Tech, 2019-06-21)
    We propose a dynamic, input dependent filter approximation and selection technique to improve the computational efficiency of Deep Neural Networks. The approximation techniques convert 32 bit floating point representation of filter weights in neural networks into smaller precision values. This is done by reducing the number of bits used to represent the weights. In order to calculate the per-input error between the trained full precision filter weights and the approximated weights, a metric called Multiplication Error (ME) has been chosen. For convolutional layers, ME is calculated by subtracting the approximated filter weights from the original filter weights, convolving the difference with the input and calculating the grand-sum of the resulting matrix. For fully connected layers, ME is calculated by subtracting the approximated filter weights from the original filter weights, performing matrix multiplication between the difference and the input and calculating the grand-sum of the resulting matrix. ME is computed to identify approximated filters in a layer that result in low inference accuracy. In order to maintain the accuracy of the network, these filters weights are replaced with the original full precision weights. Prior work has primarily focused on input independent (static) replacement of filters to low precision weights. In this technique, all the filter weights in the network are replaced by approximated filter weights. This results in a decrease in inference accuracy. The decrease in accuracy is higher for more aggressive approximation techniques. Our proposed technique aims to achieve higher inference accuracy by not approximating filters that generate high ME. Using the proposed per-input filter selection technique, LeNet achieves an accuracy of 95.6% with 3.34% drop from the original accuracy value of 98.9% for truncating to 3 bits for the MNIST dataset. On the other hand upon static filter approximation, LeNet achieves an accuracy of 90.5% with 8.5% drop from the original accuracy. The aim of our research is to potentially use low precision weights in deep learning algorithms to achieve high classification accuracy with less computational overhead. We explore various filter approximation techniques and implement a per-input filter selection and approximation technique that selects the filters to approximate during run-time.