Matching Items (259)
Filtering by
- All Subjects: Machine learning
- Creators: Arizona State University
Description
In this thesis, six experiments which were computer simulations were conducted in order to replicate the negative association between sample size and accuracy that is repeatedly found in ML literature by accounting for data leakage and publication bias. The reason why it is critical to understand why this negative association is occurring is that in published studies, there have been multiple reports that the accuracies in ML models are overoptimistic leading to cases where the results are irreproducible despite conducting multiple trials and experiments. Additionally, after replicating the negative association between sample size and accuracy, parametric curves (learning curves with the parametric function) were fitted along the empirical learning curves in order to evaluate the performance. It was found that there is a significant variance in accuracies when the sample size is small, but little to no variation when the sample size is large. In other words, the empirical learning curves with data leakage and publication bias were able to achieve the same accuracy as the learning curve without data leakage at a large sample size.
ContributorsKottooru, Rishab (Author) / Berisha, Visar (Thesis director) / Dasarathy, Gautam (Committee member) / Saidi, Pouria (Committee member) / Barrett, The Honors College (Contributor) / Chemical Engineering Program (Contributor)
Created2023-05
Description
Graph neural networks (GNN) offer a potential method of bypassing the Kohn-Sham equations in density functional theory (DFT) calculations by learning both the Hohenberg-Kohn (HK) mapping of electron density to energy, allowing for calculations of much larger atomic systems and time scales and enabling large-scale MD simulations with DFT-level accuracy. In this work, we investigate the feasibility of GNNs to learn the HK map from the external potential approximated as Gaussians to the electron density π(π), and the mapping from π(π) to the energy density π(π) using Pytorch Geometric. We develop a graph representation for densities on radial grid points and determine that a k-nearest neighbor algorithm for determining node connections is an effective approach compared to a distance cutoff model, having an average graph size of 6.31 MB and 32.0 MB for datasets with π = 10 and π = 50 respectively. Furthermore, we develop two GNNs in Pytorch Geometric, and demonstrate a decrease in training losses for a π(π) to π(π) of 8.52 Β· 10^14 and 3.10 Β· 10^14 for π = 10 and π = 20 datasets respectively, suggesting the model could be further trained and optimized to learn the electron density to energy functional.
ContributorsHayes, Matthew (Author) / Muhich, Christopher (Thesis director) / Oswald, Jay (Committee member) / Barrett, The Honors College (Contributor) / Chemical Engineering Program (Contributor) / School of Mathematical and Statistical Sciences (Contributor)
Created2023-05
Description
With the advent of Internet, the data being added online is increasing at enormous rate. Though search engines are using IR techniques to facilitate the search requests from users, the results are not effective towards the search query of the user. The search engine user has to go through certain webpages before getting at the webpage he/she wanted. This problem of Information Overload can be solved using Automatic Text Summarization. Summarization is a process of obtaining at abridged version of documents so that user can have a quick view to understand what exactly the document is about. Email threads from W3C are used in this system. Apart from common IR features like Term Frequency, Inverse Document Frequency, Term Rank, a variation of page rank based on graph model, which can cluster the words with respective to word ambiguity, is implemented. Term Rank also considers the possibility of co-occurrence of words with the corpus and evaluates the rank of the word accordingly. Sentences of email threads are ranked as per features and summaries are generated. System implemented the concept of pyramid evaluation in content selection. The system can be considered as a framework for Unsupervised Learning in text summarization.
ContributorsNadella, Sravan (Author) / Davulcu, Hasan (Thesis advisor) / Li, Baoxin (Committee member) / Sen, Arunabha (Committee member) / Arizona State University (Publisher)
Created2015
Description
Transfer learning refers to statistical machine learning methods that integrate the knowledge of one domain (source domain) and the data of another domain (target domain) in an appropriate way, in order to develop a model for the target domain that is better than a model using the data of the target domain alone. Transfer learning emerged because classic machine learning, when used to model different domains, has to take on one of two mechanical approaches. That is, it will either assume the data distributions of the different domains to be the same and thereby developing one model that fits all, or develop one model for each domain independently. Transfer learning, on the other hand, aims to mitigate the limitations of the two approaches by accounting for both the similarity and specificity of related domains. The objective of my dissertation research is to develop new transfer learning methods and demonstrate the utility of the methods in real-world applications. Specifically, in my methodological development, I focus on two different transfer learning scenarios: spatial transfer learning across different domains and temporal transfer learning along time in the same domain. Furthermore, I apply the proposed spatial transfer learning approach to modeling of degenerate biological systems.Degeneracy is a well-known characteristic, widely-existing in many biological systems, and contributes to the heterogeneity, complexity, and robustness of biological systems. In particular, I study the application of one degenerate biological system which is to use transcription factor (TF) binding sites to predict gene expression across multiple cell lines. Also, I apply the proposed temporal transfer learning approach to change detection of dynamic network data. Change detection is a classic research area in Statistical Process Control (SPC), but change detection in network data has been limited studied. I integrate the temporal transfer learning method called the Network State Space Model (NSSM) and SPC and formulate the problem of change detection from dynamic networks into a covariance monitoring problem. I demonstrate the performance of the NSSM in change detection of dynamic social networks.
ContributorsZou, Na (Author) / Li, Jing (Thesis advisor) / Baydogan, Mustafa (Committee member) / Borror, Connie (Committee member) / Montgomery, Douglas C. (Committee member) / Wu, Teresa (Committee member) / Arizona State University (Publisher)
Created2015
Description
Discriminative learning when training and test data belong to different distributions is a challenging and complex task. Often times we have very few or no labeled data from the test or target distribution, but we may have plenty of labeled data from one or multiple related sources with different distributions. Due to its capability of migrating knowledge from related domains, transfer learning has shown to be effective for cross-domain learning problems. In this dissertation, I carry out research along this direction with a particular focus on designing efficient and effective algorithms for BioImaging and Bilingual applications. Specifically, I propose deep transfer learning algorithms which combine transfer learning and deep learning to improve image annotation performance. Firstly, I propose to generate the deep features for the Drosophila embryo images via pretrained deep models and build linear classifiers on top of the deep features. Secondly, I propose to fine-tune the pretrained model with a small amount of labeled images. The time complexity and performance of deep transfer learning methodologies are investigated. Promising results have demonstrated the knowledge transfer ability of proposed deep transfer algorithms. Moreover, I propose a novel Robust Principal Component Analysis (RPCA) approach to process the noisy images in advance. In addition, I also present a two-stage re-weighting framework for general domain adaptation problems. The distribution of source domain is mapped towards the target domain in the first stage, and an adaptive learning model is proposed in the second stage to incorporate label information from the target domain if it is available. Then the proposed model is applied to tackle cross lingual spam detection problem at LinkedInβs website. Our experimental results on real data demonstrate the efficiency and effectiveness of the proposed algorithms.
ContributorsSun, Qian (Author) / Ye, Jieping (Committee member) / Xue, Guoliang (Committee member) / Liu, Huan (Committee member) / Li, Jing (Committee member) / Arizona State University (Publisher)
Created2015
Description
Robust and stable decoding of neural signals is imperative for implementing a useful neuroprosthesis capable of carrying out dexterous tasks. A nonhuman primate (NHP) was trained to perform combined flexions of the thumb, index and middle fingers in addition to individual flexions and extensions of the same digits. An array of microelectrodes was implanted in the hand area of the motor cortex of the NHP and used to record action potentials during finger movements. A Support Vector Machine (SVM) was used to classify which finger movement the NHP was making based upon action potential firing rates. The effect of four feature selection techniques, Wilcoxon signed-rank test, Relative Importance, Principal Component Analysis, and Mutual Information Maximization was compared based on SVM classification performance. SVM classification was used to examine the functional parameters of (i) efficacy (ii) endurance to simulated failure and (iii) longevity of classification. The effect of using isolated-neuron and multi-unit firing rates was compared as the feature vector supplied to the SVM. The best classification performance was on post-implantation day 36, when using multi-unit firing rates the worst classification accuracy resulted from features selected with Wilcoxon signed-rank test (51.12 Β± 0.65%) and the best classification accuracy resulted from Mutual Information Maximization (93.74 Β± 0.32%). On this day when using single-unit firing rates, the classification accuracy from the Wilcoxon signed-rank test was 88.85 Β± 0.61 % and Mutual Information Maximization was 95.60 Β± 0.52% (degrees of freedom =10, level of chance =10%)
ContributorsPadmanaban, Subash (Author) / Greger, Bradley (Thesis advisor) / Santello, Marco (Thesis advisor) / Helms Tillery, Stephen (Committee member) / Arizona State University (Publisher)
Created2015
Description
Colorectal cancer is the second-highest cause of cancer-related deaths in the United States with approximately 50,000 estimated deaths in 2015. The advanced stages of colorectal cancer has a poor five-year survival rate of 10%, whereas the diagnosis in early stages of development has showed a more favorable five-year survival rate of 90%. Early diagnosis of colorectal cancer is achievable if colorectal polyps, a possible precursor to cancer, are detected and removed before developing into malignancy.
The preferred method for polyp detection and removal is optical colonoscopy. A colonoscopic procedure consists of two phases: (1) insertion phase during which a flexible endoscope (a flexible tube with a tiny video camera at the tip) is advanced via the anus and then gradually to the end of the colon--called the cecum, and (2) withdrawal phase during which the endoscope is gradually withdrawn while colonoscopists examine the colon wall to find and remove polyps. Colonoscopy is an effective procedure and has led to a significant decline in the incidence and mortality of colon cancer. However, despite many screening and therapeutic advantages, 1 out of every 4 polyps and 1 out of 13 colon cancers are missed during colonoscopy.
There are many factors that contribute to missed polyps and cancers including poor colon preparation, inadequate navigational skills, and fatigue. Poor colon preparation results in a substantial portion of colon covered with fecal content, hindering a careful examination of the colon. Inadequate navigational skills can prevent a colonoscopist from examining hard-to-reach regions of the colon that may contain a polyp. Fatigue can manifest itself in the performance of a colonoscopist by decreasing diligence and vigilance during procedures. Lack of vigilance may prevent a colonoscopist from detecting the polyps that briefly appear in the colonoscopy videos. Lack of diligence may result in hasty examination of the colon that is likely to miss polyps and lesions.
To reduce polyp and cancer miss rates, this research presents a quality assurance system with 3 components. The first component is an automatic polyp detection system that highlights the regions with suspected polyps in colonoscopy videos. The goal is to encourage more vigilance during procedures. The suggested polyp detection system consists of several novel modules: (1) a new patch descriptor that characterizes image appearance around boundaries more accurately and more efficiently than widely-used patch descriptors such HoG, LBP, and Daisy; (2) A 2-stage classification framework that is able to enhance low level image features prior to classification. Unlike the traditional way of image classification where a single patch undergoes the processing pipeline, our system fuses the information extracted from a pair of patches for more accurate edge classification; (3) a new vote accumulation scheme that robustly localizes objects with curvy boundaries in fragmented edge maps. Our voting scheme produces a probabilistic output for each polyp candidate but unlike the existing methods (e.g., Hough transform) does not require any predefined parametric model of the object of interest; (4) and a unique three-way image representation coupled with convolutional neural networks (CNNs) for classifying the polyp candidates. Our image representation efficiently captures a variety of features such as color, texture, shape, and temporal information and significantly improves the performance of the subsequent CNNs for candidate classification. This contrasts with the exiting methods that mainly rely on a subset of the above image features for polyp detection. Furthermore, this research is the first to investigate the use of CNNs for polyp detection in colonoscopy videos.
The second component of our quality assurance system is an automatic image quality assessment for colonoscopy. The goal is to encourage more diligence during procedures by warning against hasty and low quality colon examination. We detect a low quality colon examination by identifying a number of consecutive non-informative frames in videos. We base our methodology for detecting non-informative frames on two key observations: (1) non-informative frames
most often show an unrecognizable scene with few details and blurry edges and thus their information can be locally compressed in a few Discrete Cosine Transform (DCT) coefficients; however, informative images include much more details and their information content cannot be summarized by a small subset of DCT coefficients; (2) information content is spread all over the image in the case of informative frames, whereas in non-informative frames, depending on image artifacts and degradation factors, details may appear in only a few regions. We use the former observation in designing our global features and the latter in designing our local image features. We demonstrated that the suggested new features are superior to the existing features based on wavelet and Fourier transforms.
The third component of our quality assurance system is a 3D visualization system. The goal is to provide colonoscopists with feedback about the regions of the colon that have remained unexamined during colonoscopy, thereby helping them improve their navigational skills. The suggested system is based on a new 3D reconstruction algorithm that combines depth and position information for 3D reconstruction. We propose to use a depth camera and a tracking sensor to obtain depth and position information. Our system contrasts with the existing works where the depth and position information are unreliably estimated from the colonoscopy frames. We conducted a use case experiment, demonstrating that the suggested 3D visualization system can determine the unseen regions of the navigated environment. However, due to technology limitations, we were not able to evaluate our 3D visualization system using a phantom model of the colon.
The preferred method for polyp detection and removal is optical colonoscopy. A colonoscopic procedure consists of two phases: (1) insertion phase during which a flexible endoscope (a flexible tube with a tiny video camera at the tip) is advanced via the anus and then gradually to the end of the colon--called the cecum, and (2) withdrawal phase during which the endoscope is gradually withdrawn while colonoscopists examine the colon wall to find and remove polyps. Colonoscopy is an effective procedure and has led to a significant decline in the incidence and mortality of colon cancer. However, despite many screening and therapeutic advantages, 1 out of every 4 polyps and 1 out of 13 colon cancers are missed during colonoscopy.
There are many factors that contribute to missed polyps and cancers including poor colon preparation, inadequate navigational skills, and fatigue. Poor colon preparation results in a substantial portion of colon covered with fecal content, hindering a careful examination of the colon. Inadequate navigational skills can prevent a colonoscopist from examining hard-to-reach regions of the colon that may contain a polyp. Fatigue can manifest itself in the performance of a colonoscopist by decreasing diligence and vigilance during procedures. Lack of vigilance may prevent a colonoscopist from detecting the polyps that briefly appear in the colonoscopy videos. Lack of diligence may result in hasty examination of the colon that is likely to miss polyps and lesions.
To reduce polyp and cancer miss rates, this research presents a quality assurance system with 3 components. The first component is an automatic polyp detection system that highlights the regions with suspected polyps in colonoscopy videos. The goal is to encourage more vigilance during procedures. The suggested polyp detection system consists of several novel modules: (1) a new patch descriptor that characterizes image appearance around boundaries more accurately and more efficiently than widely-used patch descriptors such HoG, LBP, and Daisy; (2) A 2-stage classification framework that is able to enhance low level image features prior to classification. Unlike the traditional way of image classification where a single patch undergoes the processing pipeline, our system fuses the information extracted from a pair of patches for more accurate edge classification; (3) a new vote accumulation scheme that robustly localizes objects with curvy boundaries in fragmented edge maps. Our voting scheme produces a probabilistic output for each polyp candidate but unlike the existing methods (e.g., Hough transform) does not require any predefined parametric model of the object of interest; (4) and a unique three-way image representation coupled with convolutional neural networks (CNNs) for classifying the polyp candidates. Our image representation efficiently captures a variety of features such as color, texture, shape, and temporal information and significantly improves the performance of the subsequent CNNs for candidate classification. This contrasts with the exiting methods that mainly rely on a subset of the above image features for polyp detection. Furthermore, this research is the first to investigate the use of CNNs for polyp detection in colonoscopy videos.
The second component of our quality assurance system is an automatic image quality assessment for colonoscopy. The goal is to encourage more diligence during procedures by warning against hasty and low quality colon examination. We detect a low quality colon examination by identifying a number of consecutive non-informative frames in videos. We base our methodology for detecting non-informative frames on two key observations: (1) non-informative frames
most often show an unrecognizable scene with few details and blurry edges and thus their information can be locally compressed in a few Discrete Cosine Transform (DCT) coefficients; however, informative images include much more details and their information content cannot be summarized by a small subset of DCT coefficients; (2) information content is spread all over the image in the case of informative frames, whereas in non-informative frames, depending on image artifacts and degradation factors, details may appear in only a few regions. We use the former observation in designing our global features and the latter in designing our local image features. We demonstrated that the suggested new features are superior to the existing features based on wavelet and Fourier transforms.
The third component of our quality assurance system is a 3D visualization system. The goal is to provide colonoscopists with feedback about the regions of the colon that have remained unexamined during colonoscopy, thereby helping them improve their navigational skills. The suggested system is based on a new 3D reconstruction algorithm that combines depth and position information for 3D reconstruction. We propose to use a depth camera and a tracking sensor to obtain depth and position information. Our system contrasts with the existing works where the depth and position information are unreliably estimated from the colonoscopy frames. We conducted a use case experiment, demonstrating that the suggested 3D visualization system can determine the unseen regions of the navigated environment. However, due to technology limitations, we were not able to evaluate our 3D visualization system using a phantom model of the colon.
ContributorsTajbakhsh, Nima (Author) / Liang, Jianming (Thesis advisor) / Greenes, Robert (Committee member) / Scotch, Matthew (Committee member) / Arizona State University (Publisher)
Created2015
Description
Feature learning and the discovery of nonlinear variation patterns in high-dimensional data is an important task in many problem domains, such as imaging, streaming data from sensors, and manufacturing. This dissertation presents several methods for learning and visualizing nonlinear variation in high-dimensional data. First, an automated method for discovering nonlinear variation patterns using deep learning autoencoders is proposed. The approach provides a functional mapping from a low-dimensional representation to the original spatially-dense data that is both interpretable and efficient with respect to preserving information. Experimental results indicate that deep learning autoencoders outperform manifold learning and principal component analysis in reproducing the original data from the learned variation sources.
A key issue in using autoencoders for nonlinear variation pattern discovery is to encourage the learning of solutions where each feature represents a unique variation source, which we define as distinct features. This problem of learning distinct features is also referred to as disentangling factors of variation in the representation learning literature. The remainder of this dissertation highlights and provides solutions for this important problem.
An alternating autoencoder training method is presented and a new measure motivated by orthogonal loadings in linear models is proposed to quantify feature distinctness in the nonlinear models. Simulated point cloud data and handwritten digit images illustrate that standard training methods for autoencoders consistently mix the true variation sources in the learned low-dimensional representation, whereas the alternating method produces solutions with more distinct patterns.
Finally, a new regularization method for learning distinct nonlinear features using autoencoders is proposed. Motivated in-part by the properties of linear solutions, a series of learning constraints are implemented via regularization penalties during stochastic gradient descent training. These include the orthogonality of tangent vectors to the manifold, the correlation between learned features, and the distributions of the learned features. This regularized learning approach yields low-dimensional representations which can be better interpreted and used to identify the true sources of variation impacting a high-dimensional feature space. Experimental results demonstrate the effectiveness of this method for nonlinear variation pattern discovery on both simulated and real data sets.
A key issue in using autoencoders for nonlinear variation pattern discovery is to encourage the learning of solutions where each feature represents a unique variation source, which we define as distinct features. This problem of learning distinct features is also referred to as disentangling factors of variation in the representation learning literature. The remainder of this dissertation highlights and provides solutions for this important problem.
An alternating autoencoder training method is presented and a new measure motivated by orthogonal loadings in linear models is proposed to quantify feature distinctness in the nonlinear models. Simulated point cloud data and handwritten digit images illustrate that standard training methods for autoencoders consistently mix the true variation sources in the learned low-dimensional representation, whereas the alternating method produces solutions with more distinct patterns.
Finally, a new regularization method for learning distinct nonlinear features using autoencoders is proposed. Motivated in-part by the properties of linear solutions, a series of learning constraints are implemented via regularization penalties during stochastic gradient descent training. These include the orthogonality of tangent vectors to the manifold, the correlation between learned features, and the distributions of the learned features. This regularized learning approach yields low-dimensional representations which can be better interpreted and used to identify the true sources of variation impacting a high-dimensional feature space. Experimental results demonstrate the effectiveness of this method for nonlinear variation pattern discovery on both simulated and real data sets.
ContributorsHoward, Phillip (Author) / Runger, George C. (Thesis advisor) / Montgomery, Douglas C. (Committee member) / Mirchandani, Pitu (Committee member) / Apley, Daniel (Committee member) / Arizona State University (Publisher)
Created2016
Description
Buildings consume nearly 50% of the total energy in the United States, which drives the need to develop high-fidelity models for building energy systems. Extensive methods and techniques have been developed, studied, and applied to building energy simulation and forecasting, while most of work have focused on developing dedicated modeling approach for generic buildings. In this study, an integrated computationally efficient and high-fidelity building energy modeling framework is proposed, with the concentration on developing a generalized modeling approach for various types of buildings. First, a number of data-driven simulation models are reviewed and assessed on various types of computationally expensive simulation problems. Motivated by the conclusion that no model outperforms others if amortized over diverse problems, a meta-learning based recommendation system for data-driven simulation modeling is proposed. To test the feasibility of the proposed framework on the building energy system, an extended application of the recommendation system for short-term building energy forecasting is deployed on various buildings. Finally, Kalman filter-based data fusion technique is incorporated into the building recommendation system for on-line energy forecasting. Data fusion enables model calibration to update the state estimation in real-time, which filters out the noise and renders more accurate energy forecast. The framework is composed of two modules: off-line model recommendation module and on-line model calibration module. Specifically, the off-line model recommendation module includes 6 widely used data-driven simulation models, which are ranked by meta-learning recommendation system for off-line energy modeling on a given building scenario. Only a selective set of building physical and operational characteristic features is needed to complete the recommendation task. The on-line calibration module effectively addresses system uncertainties, where data fusion on off-line model is applied based on system identification and Kalman filtering methods. The developed data-driven modeling framework is validated on various genres of buildings, and the experimental results demonstrate desired performance on building energy forecasting in terms of accuracy and computational efficiency. The framework could be easily implemented into building energy model predictive control (MPC), demand response (DR) analysis and real-time operation decision support systems.
ContributorsCui, Can (Author) / Wu, Teresa (Thesis advisor) / Weir, Jeffery D. (Thesis advisor) / Li, Jing (Committee member) / Fowler, John (Committee member) / Hu, Mengqi (Committee member) / Arizona State University (Publisher)
Created2016
Description
Information divergence functions, such as the Kullback-Leibler divergence or the Hellinger distance, play a critical role in statistical signal processing and information theory; however estimating them can be challenge. Most often, parametric assumptions are made about the two distributions to estimate the divergence of interest. In cases where no parametric model fits the data, non-parametric density estimation is used. In statistical signal processing applications, Gaussianity is usually assumed since closed-form expressions for common divergence measures have been derived for this family of distributions. Parametric assumptions are preferred when it is known that the data follows the model, however this is rarely the case in real-word scenarios. Non-parametric density estimators are characterized by a very large number of parameters that have to be tuned with costly cross-validation. In this dissertation we focus on a specific family of non-parametric estimators, called direct estimators, that bypass density estimation completely and directly estimate the quantity of interest from the data. We introduce a new divergence measure, the $D_p$-divergence, that can be estimated directly from samples without parametric assumptions on the distribution. We show that the $D_p$-divergence bounds the binary, cross-domain, and multi-class Bayes error rates and, in certain cases, provides provably tighter bounds than the Hellinger divergence. In addition, we also propose a new methodology that allows the experimenter to construct direct estimators for existing divergence measures or to construct new divergence measures with custom properties that are tailored to the application. To examine the practical efficacy of these new methods, we evaluate them in a statistical learning framework on a series of real-world data science problems involving speech-based monitoring of neuro-motor disorders.
ContributorsWisler, Alan (Author) / Berisha, Visar (Thesis advisor) / Spanias, Andreas (Thesis advisor) / Liss, Julie (Committee member) / Bliss, Daniel (Committee member) / Arizona State University (Publisher)
Created2017