You are here
Search results
(1 - 20 of 49)
Pages
- Title
- Robust multi-task learning algorithms for predictive modeling of spatial and temporal data
- Creator
- Liu, Xi (Graduate of Michigan State University)
- Date
- 2019
- Collection
- Electronic Theses & Dissertations
- Description
-
"Recent years have witnessed the significant growth of spatial and temporal data generated from various disciplines, including geophysical sciences, neuroscience, economics, criminology, and epidemiology. Such data have been extensively used to train spatial and temporal models that can make predictions either at multiple locations simultaneously or along multiple forecasting horizons (lead times). However, training an accurate prediction model in these domains can be challenging especially...
Show more"Recent years have witnessed the significant growth of spatial and temporal data generated from various disciplines, including geophysical sciences, neuroscience, economics, criminology, and epidemiology. Such data have been extensively used to train spatial and temporal models that can make predictions either at multiple locations simultaneously or along multiple forecasting horizons (lead times). However, training an accurate prediction model in these domains can be challenging especially when there are significant noise and missing values or limited training examples available. The goal of this thesis is to develop novel multi-task learning frameworks that can exploit the spatial and/or temporal dependencies of the data to ensure robust predictions in spite of the data quality and scarcity problems. The first framework developed in this dissertation is designed for multi-task classification of time series data. Specifically, the prediction task here is to continuously classify activities of a human subject based on the multi-modal sensor data collected in a smart home environment. As the classes exhibit strong spatial and temporal dependencies, this makes it an ideal setting for applying a multi-task learning approach. Nevertheless, since the type of sensors deployed often vary from one room (location) to another, this introduces a structured missing value problem, in which blocks of sensor data could be missing when a subject moves from one room to another. To address this challenge, a probabilistic multi-task classification framework is developed to jointly model the activity recognition tasks from all the rooms, taking into account the block-missing value problem. The framework also learns the transitional dependencies between classes to improve its overall prediction accuracy. The second framework is developed for the multi-location time series forecasting problem. Although multi-task learning has been successfully applied to many time series forecasting applications such as climate prediction, conventional approaches aim to minimize only the point-wise residual error of their predictions instead of considering how well their models fit the overall distribution of the response variable. As a result, their predicted distribution may not fully capture the true distribution of the data. In this thesis, a novel distribution-preserving multi-task learning framework is proposed for the multi-location time series forecasting problem. The framework uses a non-parametric density estimation approach to fit the distribution of the response variable and employs an L2-distance function to minimize the divergence between the predicted and true distributions. The third framework proposed in this dissertation is for the multi-step-ahead (long-range) time series prediction problem with application to ensemble forecasting of sea surface temperature. Specifically, our goal is to effectively combine the forecasts generated by various numerical models at different lead times to obtain more precise predictions. Towards this end, a multi-task deep learning framework based on a hierarchical LSTM architecture is proposed to jointly model the ensemble forecasts of different models, taking into account the temporal dependencies between forecasts at different lead times. Experiments performed on 29-year sea surface temperature data from North American Multi-Model Ensemble (NAMME) demonstrate that the proposed architecture significantly outperforms standard LSTM and other MTL approaches."--Pages ii-iii.
Show less
- Title
- Machine learning method for authorship attribution
- Creator
- Hu, Xianfeng
- Date
- 2015
- Collection
- Electronic Theses & Dissertations
- Description
-
MACHINE LEARNING METHOD FOR AUTHORSHIP ATTRIBUTIONBy Xianfeng HuBroadly speaking, the authorship identification or authorship attribution problem is to determine the authorship of a given sample such as text, painting and so on. Our main work is to develop an effective and mathe-sound approach for the analysis of authorship of doubted books.Inspired by various authorship attribution problems in the history of literature and the application of machine learning in the study of literary...
Show moreMACHINE LEARNING METHOD FOR AUTHORSHIP ATTRIBUTIONBy Xianfeng HuBroadly speaking, the authorship identification or authorship attribution problem is to determine the authorship of a given sample such as text, painting and so on. Our main work is to develop an effective and mathe-sound approach for the analysis of authorship of doubted books.Inspired by various authorship attribution problems in the history of literature and the application of machine learning in the study of literary stylometry, we develop a rigorous new method for the mathematical analysis of authorship by testing for a so-called chrono-divide in writing styles. Our method incorporates some of the latest advances in the study of au- thorship attribution, particularly techniques from support vector machines. By introducing the notion of relative frequency of word and phrases as feature ranking metrics our method proves to be highly effective and robust.Applying our method to the classical Chinese novel Dream of the Red Chamber has led to convincing if not irrefutable evidence that the first 80 chapters and the last 40 chapters of the book were written by two different authors.Also applying our method to the English novel Micro, we are able to confirm the existence of the chrono-divide and identify its location so that we can differentiate the contribution of Michael Crichton and Richard Preston, the authors of the novel.We have also tested our method to the other three Great Classical Novels in Chinese. As expected no chrono-divides have been found in these novels. This provides further evidenceof the robustness of our method. We also proposed a new approach to authorship identification to solve the open classproblem where the candidate group is nonexistent or very large, which is reliably scaled from a new method we have developed for the closed class problem in which the candidates author pool is small. This is attained by using support vector machines and by analyzing the relative frequencies of common words in the function words dictionary and most frequently used words. This method scales very nicely to the open class problem through a novel author randomization technique, where an author in question is compared repeatedly to randomly selected authors. The author randomization technique proves to be highly robust and effective. Using our approaches we have found answers to three well known authorship controversies: (1) Did Robert Galbraith write Cuckoo’s Calling? (2) Did Harper Lee write To Kill a Mockingbird or did her friend Truman Capote write it? (3) Did Bill Ayers write Obama’s autobiography Dreams From My Father?
Show less
- Title
- Image annotation and tag completion via kernel metric learning and noisy matrix recovery
- Creator
- Feng, Zheyun
- Date
- 2016
- Collection
- Electronic Theses & Dissertations
- Description
-
In the last several years, with the ever-growing popularity of digital photography and social media, the number of images with user-provided tags has increased enormously. Due to the large amount and content versatility of these images, there is an urgent need to categorize, index, retrieve and browse these images via semantic tags (also called attributes or keywords). Following this trend, image annotation or tag completion out of missing and noisy given tags over large scale datasets has...
Show moreIn the last several years, with the ever-growing popularity of digital photography and social media, the number of images with user-provided tags has increased enormously. Due to the large amount and content versatility of these images, there is an urgent need to categorize, index, retrieve and browse these images via semantic tags (also called attributes or keywords). Following this trend, image annotation or tag completion out of missing and noisy given tags over large scale datasets has become an extremely hot topic in the interdisciplinary areas of machine learning and computer vision.The overarching goal of this thesis is to reassess the image annotation and tag completion algorithms that mainly capture the essential relationship both between and within images and tags even when the given tag information is incomplete or noisy, so as to achieve a better performance in terms of both effectiveness and efficiency in image annotation and other tag relevant tasks including tag completion, tag ranking and tag refinement.One of the key challenges in search-based image annotation models is to define an appropriate similarity measure (distance metric) between images, so as to assign unlabeled images with tags that are shared among similar labeled training images. Many kernel metric learning (KML) algorithms have been developed to serve as such a nonlinear distance metric. However, most of them suffer from high computational cost since the learned kernel metric needs to be projected into a positive semi-definite (PSD) cone. Besides, in image annotation tasks, existing KML algorithms require to convert image annotation tags into binary constraints, which lead to a significant semantic information loss and severely reduces the annotation performance.In this dissertation we propose a robust kernel metric learning (RKML) algorithm based on regression technique that is able to directly utilize the image tags. RKML is computationally efficient since the PSD property is automatically ensured by the regression technique. Numeric constraints over tags are also applied to better exploit the tag information and hence improve the annotation accuracy. Further, theoretical guarantees for RKML are provided, and its efficiency and effectiveness are also verified empirically by comparing it to state-of-the-art approaches of both distance metric learning and image annotation.Since the user-provided image tags are always incomplete and noisy, we also propose a tag completion algorithm by noisy matrix recovery (TCMR) to simultaneously enrich the missing tags and remove the noisy ones. TCMR assumes that the observed tags are independently sampled from unknown distributions that are represented by a tag matrix, and our goal is to recover that tag matrix based on the partially revealed tags which could be noisy. We provide theoretical guarantees for TCMR with recovery error bounds. In addition, a graph Laplacian based component is introduced to enforce the recovered tags to be consistent with the visual contents of images. Our empirical study with multiple benchmark datasets for image tagging shows that the proposed algorithm outperforms state-of-the-art approaches in terms of both effectiveness and efficiency when handling missing and noisy tags.
Show less
- Title
- Integration of topological fingerprints and machine learning for the prediction of chemical mutagenicity
- Creator
- Cao, Yin (Quantitative analyst)
- Date
- 2017
- Collection
- Electronic Theses & Dissertations
- Description
-
"Toxicity refers to the interaction between chemical molecules that leads to adverse effects in biological systems, and mutagenicity is one of its most important endpoints. Prediction of chemical mutagenicity is essential to ensuring the safety of drugs, foods, etc. In silico modeling of chemical mutagenicity, as a replacement of in-vivo bioassays, is increasingly encouraged, due to its efficiency, effectiveness, lower cost and less reliance on animal tests. The quality of a good molecular...
Show more"Toxicity refers to the interaction between chemical molecules that leads to adverse effects in biological systems, and mutagenicity is one of its most important endpoints. Prediction of chemical mutagenicity is essential to ensuring the safety of drugs, foods, etc. In silico modeling of chemical mutagenicity, as a replacement of in-vivo bioassays, is increasingly encouraged, due to its efficiency, effectiveness, lower cost and less reliance on animal tests. The quality of a good molecular representation is usually the key to building an accurate and robust in silico model, in that each representation provides a different way for the machine to look at the molecular structure. While most molecular descriptors were introduced based on the physio-chemical and biological activities of chemical molecules, in this study, we propose a new topological representation for chemical molecules, the combinatorial topological fingerprints (CTFs) based on persistent homology, knowing that persistent homology is a suitable tool to extract global topological information from a discrete sample of points. The combination of the proposed CTFs and machine learning algorithms could give rise to efficient and powerful in silico models for mutagenic toxicity prediction. Experimental results on a developmental toxicity dataset have also shown the predictive power of the proposed CTFs and its competitive advantages of characterizing and representing chemical molecules over existing fingerprints."--Page ii.
Show less
- Title
- Smartphone-based sensing systems for data-intensive applications
- Creator
- Moazzami, Mohammad-Mahdi
- Date
- 2017
- Collection
- Electronic Theses & Dissertations
- Description
-
"Supported by advanced sensing capabilities, increasing computational resources and the advances in Artificial Intelligence, smartphones have become our virtual companions in our daily life. An average modern smartphone is capable of handling a wide range of tasks including navigation, advanced image processing, speech processing, cross app data processing and etc. The key facet that is common in all of these applications is the data intensive computation. In this dissertation we have taken...
Show more"Supported by advanced sensing capabilities, increasing computational resources and the advances in Artificial Intelligence, smartphones have become our virtual companions in our daily life. An average modern smartphone is capable of handling a wide range of tasks including navigation, advanced image processing, speech processing, cross app data processing and etc. The key facet that is common in all of these applications is the data intensive computation. In this dissertation we have taken steps towards the realization of the vision that makes the smartphone truly a platform for data intensive computations by proposing frameworks, applications and algorithmic solutions. We followed a data-driven approach to the system design. To this end, several challenges must be addressed before smartphones can be used as a system platform for data-intensive applications. The major challenge addressed in this dissertation include high power consumption, high computation cost in advance machine learning algorithms, lack of real-time functionalities, lack of embedded programming support, heterogeneity in the apps, communication interfaces and lack of customized data processing libraries. The contribution of this dissertation can be summarized as follows. We present the design, implementation and evaluation of the ORBIT framework, which represents the first system that combines the design requirements of a machine learning system and sensing system together at the same time. We ported for the first time off-the-shelf machine learning algorithms for real-time sensor data processing to smartphone devices. We highlighted how machine learning on smartphones comes with severe costs that need to be mitigated in order to make smartphones capable of real-time data-intensive processing. From application perspective we present SPOT. SPOT aims to address some of the challenges discovered in mobile-based smart-home systems. These challenges prevent us from achieving the promises of smart-homes due to heterogeneity in different aspects of smart devices and the underlining systems. We face the following major heterogeneities in building smart-homes:: (i) Diverse appliance control apps (ii) Communication interface, (iii) Programming abstraction. SPOT makes the heterogeneous characteristics of smart appliances transparent, and by that it minimizes the burden of home automation application developers and the efforts of users who would otherwise have to deal with appliance-specific apps and control interfaces. From algorithmic perspective we introduce two systems in the smartphone-based deep learning area: Deep-Crowd-Label and Deep-Partition. Deep neural models are both computationally and memory intensive, making them difficult to deploy on mobile applications with limited hardware resources. On the other hand, they are the most advanced machine learning algorithms suitable for real-time sensing applications used in the wild. Deep-Partition is an optimization-based partitioning meta-algorithm featuring a tiered architecture for smartphone and the back-end cloud. Deep-Partition provides a profile-based model partitioning allowing it to intelligently execute the Deep Learning algorithms among the tiers to minimize the smartphone power consumption by minimizing the deep models feed-forward latency. Deep-Crowd-Label is prototyped for semantically labeling user's location. It is a crowd-assisted algorithm that uses crowd-sourcing in both training and inference time. It builds deep convolutional neural models using crowd-sensed images to detect the context (label) of indoor locations. It features domain adaptation and model extension via transfer learning to efficiently build deep models for image labeling. The work presented in this dissertation covers three major facets of data-driven and compute-intensive smartphone-based systems: platforms, applications and algorithms; and helps to spurs new areas of research and opens up new directions in mobile computing research."--Pages ii-iii.
Show less
- Title
- Large-scale high dimensional distance metric learning and its application to computer vision
- Creator
- Qian, Qi
- Date
- 2015
- Collection
- Electronic Theses & Dissertations
- Description
-
Learning an appropriate distance function (i.e., similarity) is one of the key tasks in machine learning, especially for distance based machine learning algorithms, e.g., $k$-nearest neighbor classifier, $k$-means clustering, etc. Distance metric learning (DML), the subject to be studied in this dissertation, is designed to learn a metric that pulls the examples from the same class together and pushes the examples from different classes away from each other. Although many DML algorithms have...
Show moreLearning an appropriate distance function (i.e., similarity) is one of the key tasks in machine learning, especially for distance based machine learning algorithms, e.g., $k$-nearest neighbor classifier, $k$-means clustering, etc. Distance metric learning (DML), the subject to be studied in this dissertation, is designed to learn a metric that pulls the examples from the same class together and pushes the examples from different classes away from each other. Although many DML algorithms have been developed in the past decade, most of them can handle only small data sets with hundreds of features, significantly limiting their applications to real world applications that often involve millions of training examples represented by hundreds of thousands of features. Three main challenges are encountered to learn the metric from these large-scale high dimensional data: (i) To make sure that the learned metric is a Positive Semi-Definitive (PSD) matrix, a projection into the PSD cone is required at every iteration, whose cost is cubic in the dimensionality making it unsuitable for high dimensional data; (ii) The number of variables that needs to be optimized in DML is quadratic in the dimensionality, which results in the slow convergence rate in optimization and high requirement of memory storage; (iii) The number of constraints used by DML is at least quadratic, if not cubic, in the number of examples depending on if pairwise constraints or triplet constraints are used in DML. Besides, features can be redundant due to high dimensional representations (e.g., face features) and DML with feature selection is preferred for these applications.The main contribution of this dissertation is to address these challenges both theoretically and empirically. First, for the challenge arising from the PSD projection, we exploit the mini-batch strategy and adaptive sampling with smooth loss function to significantly reduce the number of updates (i.e., projections) while keeping the similar performance. Second, for the challenge arising from high dimensionality, we propose a dual random projection approach, which enjoys the light computation due to the usage of random projection and at the same time, significantly improves the effectiveness of random projection. Third, for the challenge with large-scale constraints, we develop a novel multi-stage metric learning framework. It divides the original optimization problem into multiple stages. It reduces the computation by adaptively sampling a small subset of constraints at each stage. Finally, to handle redundant features with group property, we develop a greedy algorithm that selects feature group and learns the corresponding metric simultaneously at each iteration leading to further improvement of learning efficiency when combined with adaptive mini-batch strategy and incremental sampling. Besides the theoretical and empirical investigation of DML on the benchmark datasets of machine learning, we also apply the proposed methods to several important computer vision applications (i.e., fine-grained visual categorization (FGVC) and face recognition).
Show less
- Title
- Novel learning algorithms for mining geospatial data
- Creator
- Yuan, Shuai (Software engineer)
- Date
- 2017
- Collection
- Electronic Theses & Dissertations
- Description
-
Geospatial data have a wide range of applicability in many disciplines, including environmental science, urban planning, healthcare, and public administration. The proliferation of such data in recent years have presented opportunities to develop novel data mining algorithms for modeling and extracting useful patterns from the data. However, there are many practical issues remain that must be addressed before the algorithms can be successfully applied to real-world problems. First, the...
Show moreGeospatial data have a wide range of applicability in many disciplines, including environmental science, urban planning, healthcare, and public administration. The proliferation of such data in recent years have presented opportunities to develop novel data mining algorithms for modeling and extracting useful patterns from the data. However, there are many practical issues remain that must be addressed before the algorithms can be successfully applied to real-world problems. First, the algorithms must be able to incorporate spatial relationships and other domain constraints defined by the problem. Second, the algorithms must be able to handle missing values, which are common in many geospatial data sets. In particular, the models constructed by the algorithms may need to be extrapolated to locations with no observation data. Another challenge is to adequately capture the nonlinear relationship between the predictor and response variables of the geospatial data. Accurate modeling of such relationship is not only a challenge, it is also computationally expensive. Finally, the variables may interact at different spatial scales, making it necessary to develop models that can handle multi-scale relationships present in the geospatial data. This thesis presents the novel algorithms I have developed to overcome the practical challenges of applying data mining to geospatial datasets. Specifically, the algorithms will be applied to both supervised and unsupervised learning problems such as cluster analysis and spatial prediction. While the algorithms are mostly evaluated on datasets from the ecology domain, they are generally applicable to other geospatial datasets with similar characteristics. First, a spatially constrained spectral clustering algorithm is developed for geospatial data. The algorithm provides a flexible way to incorporate spatial constraints into the spectral clustering formulation in order to create regions that are spatially contiguous and homogeneous. It can also be extended to a hierarchical clustering setting, enabling the creation of fine-scale regions that are nested wholly within broader-scale regions. Experimental results suggest that the nested regions created using the proposed approach are more balanced in terms of their sizes compared to the regions found using traditional hierarchical clustering methods. Second, a supervised hash-based feature learning algorithm is proposed for modeling nonlinear relationships in incomplete geospatial data. The proposed algorithm can simultaneously infer missing values while learning a small set of discriminative, nonlinear features of the geospatial data. The efficacy of the algorithm is demonstrated using synthetic and real-world datasets. Empirical results show that the algorithm is more effective than the standard approach of imputing the missing values before applying nonlinear feature learning in more than 75% of the datasets evaluated in the study. Third, a multi-task learning framework is developed for modeling multiple response variables in geospatial data. Instead of training the local models independently for each response variable at each location, the framework simultaneously fits the local models for all response variables by optimizing a joint objective function with trace-norm regularization. The framework also leverages the spatial autocorrelation between locations as well as the inherent correlation between response variables to improve prediction accuracy. Finally, a multi-level, multi-task learning framework is proposed to effectively train predictive models from nested geospatial data containing predictor variables measured at multiple spatial scales. The framework enables distinct models to be developed for each coarse- scale region using both its fine-level and coarse-level features. It also allows information to be shared among the models through a common set of latent features. Empirical results show that such information sharing helps to create more robust models especially for regions with limited or no training data. Another advantage of using the multi-level, multi-task learning framework is that it can automatically identify potential cross-scale interactions between the regional and local variables.
Show less
- Title
- Machine learning for the study of gene regulation and complex traits
- Creator
- Sonnenschein, Anne
- Date
- 2017
- Collection
- Electronic Theses & Dissertations
- Description
-
"Functional elements are found in DNA outside of protein coding regions; an important class of these elements are 'enhancers', which govern when and where transcription occurs. Predicting the identity and function of potential enhancers based on DNA sequence remains a major goal of genomics. A number of features are associated with the enhancer state, but even combinations of these features in well-studied systems such as Drosophila have limited predictive accuracy. I have examined the...
Show more"Functional elements are found in DNA outside of protein coding regions; an important class of these elements are 'enhancers', which govern when and where transcription occurs. Predicting the identity and function of potential enhancers based on DNA sequence remains a major goal of genomics. A number of features are associated with the enhancer state, but even combinations of these features in well-studied systems such as Drosophila have limited predictive accuracy. I have examined the current limits of computational enhancer prediction, and analyzed which features are most useful for this task, by applying machine-learning methods to an extensive set of genomic features. Inferring the genetic underpinning of even well-characterized phenotypes is equally challenging, although similar analytical methods can be applied. Phenotypes are frequently defined based on a set of characteristic features; when images are used as specimens, these features are frequently based on morphometric landmarks, although computational pattern-recognition has been used as an alternative. I use Drosophila wing shape as a model for a complex phenotype, and use machine learning to predict underlying genotype using both traditional landmarks and features extracted using 'computer vision."--Page ii.
Show less
- Title
- Using developmental learning for network intrusion detection
- Creator
- Knoester, David B.
- Date
- 2004
- Collection
- Electronic Theses & Dissertations
- Title
- Towards a learning system for robot hand-eye coordination
- Creator
- Howden, Sally Jean
- Date
- 1996
- Collection
- Electronic Theses & Dissertations
- Title
- Signal processing and machine learning approaches to enabling advanced sensing and networking capabilities in everyday infrastructure and electronics
- Creator
- Ali, Kamran (Scientist)
- Date
- 2020
- Collection
- Electronic Theses & Dissertations
- Description
-
Mainstream commercial off-the-shelf (COTS) electronic devices of daily use are usually designed and manufactured to serve a very specific purpose. For example, the WiFi routers and network interface cards (NICs) are designed for high speed wireless communication, RFID readers and tags are designed to identify and track items in supply chain, and smartphone vibrator motors are designed to provide haptic feedback (e.g. notifications in silent mode) to the users. This dissertation focuses on...
Show moreMainstream commercial off-the-shelf (COTS) electronic devices of daily use are usually designed and manufactured to serve a very specific purpose. For example, the WiFi routers and network interface cards (NICs) are designed for high speed wireless communication, RFID readers and tags are designed to identify and track items in supply chain, and smartphone vibrator motors are designed to provide haptic feedback (e.g. notifications in silent mode) to the users. This dissertation focuses on revisiting the physical-layer of various such everyday COTS electronic devices, either to leverage the signals obtained from their physical layers to develop novel sensing applications, or to modify/improve their PHY/MAC layer protocols to enable even more useful deployment scenarios and networking applications - while keeping their original purpose intact - by introducing mere software/firmware level changes and completely avoiding any hardware level changes. Adding such new usefulness and functionalities to existing everyday infrastructure and electronics has advantages both in terms of cost and convenience of use/deployment, as those devices (and their protocols) are already mainstream, easily available, and often already purchased and in use/deployed to serve their mainstream purpose of use.In our works on WiFi signals based sensing, we propose signal processing and machine learning approaches to enable fine-grained gesture recognition and sleep monitoring using COTS WiFi devices. In our work on gesture recognition, we show for the first time thatWiFi signals can be used to recognize small gestures with high accuracy. In our work on sleep monitoring, we propose for the first time aWiFi CSI based sleep quality monitoring scheme which can robustly track breathing and body/limb activity related vital signs during sleep throughout a night in an individual and environment independent manner.In our work on RFID signals based sensing, we propose signal processing and machine learning approaches to effectively image customer activity in front of display items in places such as retail stores using commercial off-the-shelf (COTS) monostatic RFID devices (i.e. which use a single antenna at a time for both transmitting and receiving RFID signals to and from the tags). The key novelty of this work is on achieving multi-person activity tracking in front of display items by constructing coarse grained images via robust, analytical model-driven deep learning based, RFID imaging. We implemented our scheme using a COTS RFID reader and tags.In our work on smartphone's vibration based sensing, we propose a robust and practical vibration based sensing scheme that works with smartphones with different hardware, can extract fine-grained vibration signatures of different surfaces, and is robust to environmental noise and hardware based irregularities. A useful application of this sensing is symbolic localization/tagging, e.g. figuring out whether a user's device is in their hand, pocket, or at their bedroom table, etc. Such symbolic tagging of locations can provide us with indirect information about user activities and intentions without any dedicated infrastructure, based on which we can enable useful services such as context aware notifications/alarms. To make our scheme easily scalable and compatible with COTS smartphones, we design our signal processing and machine learning pipeline such that it relies only on builtin vibration motors and microphone for sensing, and it is robust to hardware irregularities and background environmental noises. We tested our scheme on two different Android smartphones.In our work on powerline communications (PLCs), we propose a distributed spectrum sharing scheme for enterprise level PLC mesh networks. This work is a major step towards using existing COTS PLC devices to connect different types of Internet of Things (IoT) devices for sensing and control related applications in large campuses such as enterprises. Our work is based on identification of a key weakness of the existing HomePlug AV (HPAV) PLC protocol that it does not support spectrum sharing, i.e., currently each link operates over the whole available spectrum, and therefore, only one link can operate at a time. Our proposed spectrum sharing scheme significantly boosts both aggregated and per-link throughputs, by allowing multiple links to communicate concurrently, while requiring a few modifications to the existing HPAV protocol.
Show less
- Title
- Online innovization : towards knowledge discovery and achieving faster convergence in multi-objective optimization
- Creator
- Gaur, Abhinav
- Date
- 2020
- Collection
- Electronic Theses & Dissertations
- Description
-
Ì0300nnovization'' is a task of learning common principles thatexist among some or all of the Pareto-optimal solutions in amulti-objective optimization problem. Except a few earlierstudies, most innovization related studies were performed onthe final non-dominated solutions found by an evolutionary multi-objective algorithm eithermanually or by using a machine learning method.Recent studies have shown that these principles can be learnedduring intermediate iterations of an optimization run...
Show moreÌ0300nnovization'' is a task of learning common principles thatexist among some or all of the Pareto-optimal solutions in amulti-objective optimization problem. Except a few earlierstudies, most innovization related studies were performed onthe final non-dominated solutions found by an evolutionary multi-objective algorithm eithermanually or by using a machine learning method.Recent studies have shown that these principles can be learnedduring intermediate iterations of an optimization run and simultaneously utilized in thesame optimization run to repair variables to achieve a fasterconvergence to the Pareto-optimal set. This is what we are calling as ò0300nline innovization'' as it is performed online during the run of an evolutionary multi-objective optimization algorithm. Special attention is paid to learning rules that are easier to interpret, such as short algebraic expressions, instead of complex decision trees or kernel based black box rules.We begin by showing how to learn fixed form rules that are encountered frequently in multi-objective optimization problems. We also show how can we learn free form rules, that are linear combination of non-linear terms, using a custom genetic programming algorithm. We show how can we use the concept of k0300nee' in PO set of solutions along with a custom dimensional penalty calculator to discard rules that may be overly complex, or inaccurate or just dimensionally incorrect. The results of rules learned using this custom genetic programming algorithm show that it is beneficial to let evolution learn the structure of rules while the constituent weights should be learned using some classical learning algorithm such as linear regression or linear support vector machines. When the rules are implicit functions of the problem variables, we use a computationally inexpensive way of repairing the variables by turning the problem of repairing the variable into a single variable golden section search.We show the proof of concept on test problems by learning fixed form rules among variables of the problem, which we then use during the same optimization run to repair variables. Different principleslearned during an optimization run can involve differentnumber of variables and/or variables that arecommon among a number of principles. Moreover, a preferenceorder for repairing variables may play an important role forproper convergence. Thus, when multiple principles exist, itis important to use a strategy that is most beneficial forrepairing evolving population of solutions.The above methods are applied to a mix of test problems and engineering design problems. The results are encouraging and strongly supportsthe use of innovization task in enhancing the convergence of an evolutionary multi-objective optimization algorithms. Moreover, the custom genetic program developed in this work can be a useful machine learning tool for practitioners to learn human interpretable rules in the form of algebraic expressions.
Show less
- Title
- Efficient extended Kalman filter learning for feedforward layered neural networks
- Creator
- Benromdhane, Saida
- Date
- 1996
- Collection
- Electronic Theses & Dissertations
- Title
- Learning 3D model from 2D in-the-wild images
- Creator
- Tran, Luan Quoc
- Date
- 2020
- Collection
- Electronic Theses & Dissertations
- Description
-
Understanding 3D world is one of computer vision's fundamental problems. While a human has no difficulty understanding the 3D structure of an object upon seeing its 2D image, such a 3D inferring task remains extremely challenging for computer vision systems. To better handle the ambiguity in this inverse problem, one must rely on additional prior assumptions such as constraining faces to lie in a restricted subspace from a 3D model. Conventional 3D models are learned from a set of 3D scans or...
Show moreUnderstanding 3D world is one of computer vision's fundamental problems. While a human has no difficulty understanding the 3D structure of an object upon seeing its 2D image, such a 3D inferring task remains extremely challenging for computer vision systems. To better handle the ambiguity in this inverse problem, one must rely on additional prior assumptions such as constraining faces to lie in a restricted subspace from a 3D model. Conventional 3D models are learned from a set of 3D scans or computer-aided design (CAD) models, and represented by two sets of PCA basis functions. Due to the type and amount of training data, as well as, the linear bases, the representation power of these model can be limited. To address these problems, this thesis proposes an innovative framework to learn a nonlinear 3D model from a large collection of in-the-wild images, without collecting 3D scans. Specifically, given an input image (of a face or an object), a network encoder estimates the projection, lighting, shape and albedo parameters. Two decoders serve as the nonlinear model to map from the shape and albedo parameters to the 3D shape and albedo, respectively. With the projection parameter, lighting, 3D shape, and albedo, a novel analytically differentiable rendering layer is designed to reconstruct the original input. The entire network is end-to-end trainable with only weak supervision. We demonstrate the superior representation power of our models on different domains (face, generic objects), and their contribution to many other applications on facial analysis and monocular 3D object reconstruction.
Show less
- Title
- Machine learning for pose selection
- Creator
- Pei, Jun (Graduate of Michigan State University)
- Date
- 2020
- Collection
- Electronic Theses & Dissertations
- Description
-
Scoring functions play an important role in protein related systems. In general, scoring functions were developed to connect three dimensional structures and corresponding stabilities. In protein- folding systems, scoring functions can be used to predict the most stable protein structure; in protein-ligand and protein-protein systems, scoring functions can be used to find the best ligand structure, predict the binding affinities, and identifying the correct binding modes. Potential functions...
Show moreScoring functions play an important role in protein related systems. In general, scoring functions were developed to connect three dimensional structures and corresponding stabilities. In protein- folding systems, scoring functions can be used to predict the most stable protein structure; in protein-ligand and protein-protein systems, scoring functions can be used to find the best ligand structure, predict the binding affinities, and identifying the correct binding modes. Potential functions make up an essential part of scoring functions. Each potential function usually represents a different interaction that exists in a protein or protein-ligand system. In many traditional scoring functions, energies calculated from individual potential functions were simply sum up to estimate the stability of the whole structure. However, it is possible that those energies cannot be directly added together. In other words, some of those potential functions might describe more important interactions, whereas other potential functions are used to represent insignificant interactions. Hence, it will be useful to construct a model, which can emphasize the important interactions, andignore the insignificant ones.With the development of machine learning (ML), it became possible to build up a model, which can address the importance of different interactions. In this work, we combined random forest (RF) algorithm and different potential function sets to solve the pose selection problem in protein- folding and protein-ligand systems. Chapter 3 and chapter 5 show the results of combing RF algorithm with knowledge-based potential functions and force field potential functions for protein-folding systems. Chapter 4 shows the result of combining the RF method with knowledge-based potential functions for protein-ligand systems. As the results from chapter 3, chapter 4, and chapter 5, it is obvious that the RF model based on potential functions outperformed all of the traditional scoring functions in accuracy and native ranking tests. In order to test the importance of potential functions, scrambled and uniform artificial potential function sets were generated in chapter 3, the test results suggest that the potential function set is important in the model, and the most useful information from knowledge-base potential functions are the peak positions. In chapter 5, the importance of the RF algorithm and potential functions were tested. The results also suggest that the potential functions are important, and the RF model is also necessary to achieve the best performance.
Show less
- Title
- Collaborative learning : theory, algorithms, and applications
- Creator
- Lin, Kaixiang
- Date
- 2020
- Collection
- Electronic Theses & Dissertations
- Description
-
Human intelligence prospers with the advantage of collaboration. To solve one or a set of challenging tasks, we can effectively interact with peers, fuse knowledge from different sources, continuously inspire, contribute, and develop the expertise for the benefit of the shared objectives. Human collaboration is flexible, adaptive, and scalable in terms of various cooperative constructions, collaborating across interdisciplinary, even seemingly unrelated domains, and building large-scale...
Show moreHuman intelligence prospers with the advantage of collaboration. To solve one or a set of challenging tasks, we can effectively interact with peers, fuse knowledge from different sources, continuously inspire, contribute, and develop the expertise for the benefit of the shared objectives. Human collaboration is flexible, adaptive, and scalable in terms of various cooperative constructions, collaborating across interdisciplinary, even seemingly unrelated domains, and building large-scale disciplined organizations for extremely complex tasks. On the other hand, while machine intelligence achieved tremendous success in the past decade, the ability to collaboratively solve complicated tasks is still limited compared to human intelligence.In this dissertation, we study the problem of collaborative learning - building flexible, generalizable, and scalable collaborative strategies to facilitate the efficiency of learning one or a set of objectives. Towards achieving this goal, we investigate the following concrete and fundamental problems:1. In the context of multi-task learning, can we enforce flexible forms of interactions from multiple tasks and adaptively incorporate human expert knowledge to guide the collaboration?2. In reinforcement learning, can we design collaborative methods that effectivelycollaborate among heterogeneous learning agents to improve the sample-efficiency?3. In multi-agent learning, can we develop a scalable collaborative strategy to coordinate a massive number of learning agents accomplishing a shared task?4. In federated learning, can we have provable benefit from increasing the number of collaborative learning agents?This thesis provides the first line of research to view the above learning fields in a unified framework, which includes novel algorithms for flexible, adaptive collaboration, real-world applications using scalable collaborative learning solutions, and fundamental theories for propelling the understanding of collaborative learning.
Show less
- Title
- Algebraic topology and machine learning for biomolecular modeling
- Creator
- Cang, Zixuan
- Date
- 2018
- Collection
- Electronic Theses & Dissertations
- Description
-
Data is expanding in an unprecedented speed in both quantity and size. Topological data analysis provides excellent tools for analyzing high dimensional and highly complex data. Inspired by the topological data analysis's ability of robust and multiscale characterization of data and motivated by the demand of practical predictive tools in computational biology and biomedical researches, this dissertation extends the capability of persistent homology toward quantitative and predictive data...
Show moreData is expanding in an unprecedented speed in both quantity and size. Topological data analysis provides excellent tools for analyzing high dimensional and highly complex data. Inspired by the topological data analysis's ability of robust and multiscale characterization of data and motivated by the demand of practical predictive tools in computational biology and biomedical researches, this dissertation extends the capability of persistent homology toward quantitative and predictive data analysis tools with an emphasis in biomolecular systems. Although persistent homology is almost parameter free, careful treatment is still needed toward practically useful prediction models for realistic systems. This dissertation carefully assesses the representability of persistent homology for biomolecular systems and introduces a collection of characterization tools for both macromolecules and small molecules focusing on intra- and inter-molecular interactions, chemical complexities, electrostatics, and geometry. The representations are then coupled with deep learning and machine learning methods for several problems in drug design and biophysical research. In real-world applications, data often come with heterogeneous dimensions and components. For example, in addition to location, atoms of biomolecules can also be labeled with chemical types, partial charges, and atomic radii. While persistent homology is powerful in analyzing geometry of data, it lacks the ability of handling the non-geometric information. Based on cohomology, we introduce a method that attaches the non-geometric information to the topological invariants in persistent homology analysis. This method is not only useful to handle biomolecules but also can be applied to general situations where the data carries both geometric and non-geometric information. In addition to describing biomolecular systems as a static frame, we are often interested in the dynamics of the systems. An efficient way is to assign an oscillator to each atom and study the coupled dynamical system induced by atomic interactions. To this end, we propose a persistent homology based method for the analysis of the resulting trajectories from the coupled dynamical system. The methods developed in this dissertation have been applied to several problems, namely, prediction of protein stability change upon mutations, protein-ligand binding affinity prediction, virtual screening, and protein flexibility analysis. The tools have shown top performance in both commonly used validation benchmarks and community-wide blind prediction challenges in drug design.
Show less
- Title
- Integration of topological data analysis and machine learning for small molecule property predictions
- Creator
- Wu, Kedi
- Date
- 2018
- Collection
- Electronic Theses & Dissertations
- Description
-
Accurate prediction of small molecule properties is of paramount importance to drug design and discovery. A variety of quantitative properties of small molecules has been studied in this thesis. These properties include solvation free energy, partition coefficient, aqueous solubility, and toxicity endpoints. The highlight of this thesis is to introduce an algebraic topology based method, called element specific persistent homology (ESPH), to predict small molecule properties. Essentially ESPH...
Show moreAccurate prediction of small molecule properties is of paramount importance to drug design and discovery. A variety of quantitative properties of small molecules has been studied in this thesis. These properties include solvation free energy, partition coefficient, aqueous solubility, and toxicity endpoints. The highlight of this thesis is to introduce an algebraic topology based method, called element specific persistent homology (ESPH), to predict small molecule properties. Essentially ESPH describes molecular properties in terms of multiscale and multicomponent topological invariants and is different from conventional chemical and physical representations. Based on ESPH and its modified version, element-specific topological descriptors (ESTDs) are constructed. The advantage of ESTDs is that they are systematical, comprehensive, and scalable with respect to molecular size and composition variations, and are readily suitable for machine learning methods, rendering topological learning algorithms. Due to the inherent correlation between different small molecule properties, multi-task frameworks are further employed to simultaneously predict related properties. Deep neural networks, along with ensemble methods such as random forest and gradient boosting trees, are used to develop quantitative predictive models. Physical based molecular descriptors and auxiliary descriptors are also used in addition to ESTDs. As a result, we obtain state-of-the-art results for various benchmark data sets of small molecule properties. We have also developed two online servers for predicting properties of small molecules, TopP-S and TopTox. TopP-S is a software for topological learning predictions of partition coefficient and aqueous solubility, and TopTox is a software for computing element-specific tological descriptors (ESTDs) for toxicity endpoint predictions. They are available at http://weilab.math.msu.edu/TopP-S/ and http://weilab.math.msu.edu/TopTox/, respectively.
Show less
- Title
- Adaptive on-device deep learning systems
- Creator
- Fang, Biyi
- Date
- 2019
- Collection
- Electronic Theses & Dissertations
- Description
-
"Mobile systems such as smartphones, drones, and augmented-reality headsets are revolutionizing our lives. On-device deep learning is regarded as the key enabling technology for realizing their full potential. This is because communication with cloud adds additional latency or cost, or the applications must operate even with intermittent internet connectivity.The key to achieving the full promise of these mobile vision systems is effectively analyzing the streaming video frames. However,...
Show more"Mobile systems such as smartphones, drones, and augmented-reality headsets are revolutionizing our lives. On-device deep learning is regarded as the key enabling technology for realizing their full potential. This is because communication with cloud adds additional latency or cost, or the applications must operate even with intermittent internet connectivity.The key to achieving the full promise of these mobile vision systems is effectively analyzing the streaming video frames. However, processing streaming video frames taken in mobile settings is challenging in two folds. First, the processing usually involves multiple computer vision tasks. This multi-tenant characteristic requires mobile vision systems to concurrently run multiple applications that target different vision tasks. Second, the context in mobile settings can be frequently changed. This requires mobile vision systems to be able to switch applications to execute new vision tasks encountered in the new context.In this article, we fill this critical gap by proposing NestDNN, a framework that enables resource-aware multi-tenant on-device deep learning for continuous mobile vision. NestDNN enables each deep learning model to offer flexible resource-accuracy trade-offs. At runtime,it dynamically selects the optimal resource-accuracy trade-off for each deep learning model to fit the model's resource demand to the system's available runtime resources. In doing so, NestDNN efficiently utilizes the limited resources in mobile vision systems to jointly maximize the performance of all the concurrently running applications.Although NestDNN is able to efficiently utilize the resource by being resource-aware, it essentially treats the content of each input image equally and hence does not realize the full potential of such pipelines. To realize its full potential, we further propose FlexDNN, a novel content-adaptive framework that enables computation-efficient DNN-based on-device video stream analytics based on early exit mechanism. Compared to state-of-the-art earlyexit-based solutions, FlexDNN addresses their key limitations and pushes the state-of-the-artforward through its innovative fine-grained design and automatic approach for generating the optimal network architecture."--Pages ii-iii.
Show less
- Title
- A systematic evaluation of computational models of phonotactics
- Creator
- Sarver, Isaac
- Date
- 2020
- Collection
- Electronic Theses & Dissertations
- Description
-
In this thesis, recent computational models of phonotactics are discussed and evaluated and two new models are implemented. Prior phonotactic modeling, motivated by gradient acceptability judgments in nonce word judgment tasks (Albright 2009), claim that phonotactic grammaticality is gradient, and these models are evaluated by their ability to judge nonce words with scores that correlate with human acceptability judgments. Gorman (2013) argues that these gradient models do not account for the...
Show moreIn this thesis, recent computational models of phonotactics are discussed and evaluated and two new models are implemented. Prior phonotactic modeling, motivated by gradient acceptability judgments in nonce word judgment tasks (Albright 2009), claim that phonotactic grammaticality is gradient, and these models are evaluated by their ability to judge nonce words with scores that correlate with human acceptability judgments. Gorman (2013) argues that these gradient models do not account for the facts sufficiently and claims phonotactic grammaticality is categorical. In this thesis, the account of Gorman (2013) is implemented as well as a prominent gradient model from Hayes and Wilson (2008) and compared with the performance of two machine learning models (a support vector machine and a recurrent neural network), with all models trained on a corpus of English onsets. Results in this thesis show that the computational models are unable to correlate with human judgment data from Scholes (1966) as well as a categorical prediction of acceptability based on whether a sequence is attested in the lexicon or not, and that these models rely on assumptions which when challenged show that the models do not convincingly capture the gradience of the human judgment data used for evaluation.
Show less