skip to main content
OSTI.GOV title logo U.S. Department of Energy
Office of Scientific and Technical Information

Title: Analysis of Pre-Trained Deep Neural Networks for Large-Vocabulary Automatic Speech Recognition

Authors:
Publication Date:
Research Org.:
Lawrence Livermore National Lab. (LLNL), Livermore, CA (United States)
Sponsoring Org.:
USDOE
OSTI Identifier:
1289367
Report Number(s):
LLNL-TH-698797
DOE Contract Number:
AC52-07NA27344
Resource Type:
Thesis/Dissertation
Country of Publication:
United States
Language:
English
Subject:
97 MATHEMATICS, COMPUTING, AND INFORMATION SCIENCE

Citation Formats

Stevenson, G A. Analysis of Pre-Trained Deep Neural Networks for Large-Vocabulary Automatic Speech Recognition. United States: N. p., 2016. Web. doi:10.2172/1289367.
Stevenson, G A. Analysis of Pre-Trained Deep Neural Networks for Large-Vocabulary Automatic Speech Recognition. United States. doi:10.2172/1289367.
Stevenson, G A. 2016. "Analysis of Pre-Trained Deep Neural Networks for Large-Vocabulary Automatic Speech Recognition". United States. doi:10.2172/1289367. https://www.osti.gov/servlets/purl/1289367.
@article{osti_1289367,
title = {Analysis of Pre-Trained Deep Neural Networks for Large-Vocabulary Automatic Speech Recognition},
author = {Stevenson, G A},
abstractNote = {},
doi = {10.2172/1289367},
journal = {},
number = ,
volume = ,
place = {United States},
year = 2016,
month = 7
}

Thesis/Dissertation:
Other availability
Please see Document Availability for additional information on obtaining the full-text document. Library patrons may search WorldCat to identify libraries that hold this thesis or dissertation.

Save / Share:
  • The Cryogenic Dark Matter Search (CDMS) experiment is designed to search for dark matter in the form of Weakly Interacting Massive Particles (WIMPs) via their elastic scattering interactions with nuclei. This dissertation presents the CDMS detector technology and the commissioning of two towers of detectors at the deep underground site in Soudan, Minnesota. CDMS detectors comprise crystals of Ge and Si at temperatures of 20 mK which provide ~keV energy resolution and the ability to perform particle identification on an event by event basis. Event identification is performed via a two-fold interaction signature; an ionization response and an athermal phononmore » response. Phonons and charged particles result in electron recoils in the crystal, while neutrons and WIMPs result in nuclear recoils. Since the ionization response is quenched by a factor ~ 3(2) in Ge(Si) for nuclear recoils compared to electron recoils, the relative amplitude of the two detector responses allows discrimination between recoil types. The primary source of background events in CDMS arises from electron recoils in the outer 50 µm of the detector surface which have a reduced ionization response. We develop a quantitative model of this ‘dead layer’ effect and successfully apply the model to Monte Carlo simulation of CDMS calibration data. Analysis of data from the two tower run March-August 2004 is performed, resulting in the world’s most sensitive limits on the spin-independent WIMP-nucleon cross-section, with a 90% C.L. upper limit of 1.6 × 10 -43 cm 2 on Ge for a 60 GeV WIMP. An approach to performing surface event discrimination using neural networks and wavelets is developed. A Bayesian methodology to classifying surface events using neural networks is found to provide an optimized method based on minimization of the expected dark matter limit. The discrete wavelet analysis of CDMS phonon pulses improves surface event discrimination in conjunction with the neural network analysis, giving a 20% improvement to the expected and final WIMP-nucleon cross-section upper limits.« less
  • Artificial neural networks provide a new approach to commodity forecasting that does not require algorithm or rule development. Neural networks have been deemed successful in applications involving optimization, classification, identification, pattern recognition and time series forecasting. With the advent of user friendly, commercially available software packages that work in a spreadsheet environment, such as Neural Works Predict by NeuralWare, more people can take advantage of the power of artificial neural networks. This thesis provides an introduction to neural networks, and reviews two recent studies of forecasting commodities prices. This study also develops a neural network model using Neural Works Predictmore » that forecasts jet fuel prices for the Defense Fuel Supply Center (DFSC). In addition, the results developed are compared to the output of an econometric regression model, specifically, the Department of Energy`s Short-Term Integrated Forecasting System (STWS) model. The Predict artificial neural network model produced more accurate results and reduced the contribution of outliers more effectively than the STIFS model, thus producing a more robust model.« less
  • The authors present the results of a new measurement of the tmore » $$\bar{t}$$ production cross section using eμ channel in p$$\bar{p}$$ collisions at √s = 1.8 TeV. This study corresponds to an integrated luminosity of 108.3 ± 5.7 pb -1 acquired by the D0 detector during the Fermilab Tevatron Collider Run 1 (1992--1996). By using neural network techniques instead of the conventional analysis methods, the authors show that the signal acceptance can be increased by 10% (for m t = 172 GeV/c 2) while the background remains constant. Four eμ events are observed in data with an estimated background of 0.22 ± 0.14 corresponding to a t$$\bar{t}$$ production cross section of 9.75 ± 5.53 pb.« less
  • We present results of a search for single top quark production in pmore » $$\bar{p}$$ collisions using a dataset of approximately 1 fb -1 collected with the D0 detector. This analysis considers the muon+jets and electron+jets final states and makes use of Bayesian neural networks to separate the expected signals from backgrounds. The observed excess is associated with a p-value of 0.081%, assuming the background-only hypothesis, which corresponds to an excess over background of 3.2 standard deviations for a Gaussian density. The p-value computed using the SM signal cross section of 2.9 pb is 1.6%, corresponding to an expected significance of 2.2 standard deviations. Assuming the observed excess is due to single top production, we measure a single top quark production cross section of σ(p$$\bar{p}$$ → tb + X, tqb + X) = 4.4 ± 1.5 pb.« less
  • CDF (Collider Detector at Fermilab) is a particle detector located at Fermi National Laboratories, near Chicago. it allows to study decay products of pmore » $$\bar{p}$$ collisions at center-of-mass energy of 1.96 TeV. During its first period of data taking (RunI), CDF observed for the first time the top quark (1995). The current period of data taking (RunII) is devoted to precise measurements of top properties and to search for new physics. This thesis work is about the top decay channel named τ + jets. A t$$\bar{t}$$ pair decays in two W bosons and two b quarks. In a τ + jets event, one out of the two W decays into two jets of hadrons, while the other produces a τ lepton and a neutrino; the τ decays semileptonically in one or more charged and neutral pions while b quarks hadronize producing two jets of particles. Thus the final state of a τ + jets event has this specific signature: five jets, one τ-like, i.e. narrow and with low track multiplicity, two from b quarks, two from a W boson and a large amount of missing energy from two τ neutrinos. They search for this signal in 311 pb -1 of data collected with TOP{_}MULTIJET trigger. They use neural networks to separate signal from background and on the selected sample they perform a t$$\bar{t}$$ production cross section measurement. The thesis is structured as follows: in Chapter 1 they outline the physics of top and τ, concentrating on their discovery, production mechanisms and current physics results involving them. Chapter 2 is devoted to the description of the experimental setup: the accelerator complex first and CDF detector then. The trigger system is described in Chapter 3, while Chapter 4 shows how particles are reconstructed exploiting information from different CDF subdetectors. With Chapter 5 they begin to present their analysis: we use a feed forward neural network based on a minimization algorithm developed in Trento University, called Reactive Taboo Search (RTS), especially designed to rapidly escape from local minima. Using this neural network, they explore two techniques to select t$$\bar{t}$$ → τ + jets events, the first based on a single net, the second on two neural networks in cascade; both techniques are described in Chapter 6, together with the variables used as inputs for the nets. Finally, in Chapter 7 they present a method to measure cross section on the sample of events selected by neural networks.« less