Dissertations / Theses on the topic '170203 Knowledge Representation and Machine Learning'

Recent progress in machine learning has been mainly due to
the availability of large amounts of annotated data used for training complex
models with deep architectures. Annotating this training data becomes
burdensome and creates a major bottleneck in maintaining machine-learning
databases. Moreover, these trained models fail to generalize to new categories
or new varieties of the same categories. This is because new categories or new
varieties have data distribution different from the training data distribution.
To tackle these problems, this thesis proposes to develop a family of
transfer-learning techniques that can deal with different training (source) and
testing (target) distributions with the assumption that the availability of
annotated data is limited in the testing domain. This is done by using the
auxiliary data-abundant source domain from which useful knowledge is
transferred that can be applied to data-scarce target domain. This transferable
knowledge serves as a prior that biases target-domain predictions and prevents
the target-domain model from overfitting. Specifically, we explore structural
priors that encode relational knowledge between different data entities, which
provides more informative bias than traditional priors. The choice of the
structural prior depends on the information availability and the similarity
between the two domains. Depending on the domain similarity and the information
availability, we divide the transfer learning problem into four major
categories and propose different structural priors to solve each of these
sub-problems.


This thesis first focuses on the
unsupervised-domain-adaptation problem, where we propose to minimize domain
discrepancy by transforming labeled source-domain data to be close to unlabeled
target-domain data.  For this problem,
the categories remain the same across the two domains and hence we assume that
the structural relationship between the source-domain samples is carried over
to the target domain. Thus, graph or hyper-graph is constructed as the
structural prior from both domains and a graph/hyper-graph matching formulation
is used to transform samples in the source domain to be closer to samples in
the target domain. An efficient optimization scheme is then proposed to tackle
the time and memory inefficiencies associated with the matching problem. The
few-shot learning problem is studied next, where we propose to transfer
knowledge from source-domain categories containing abundantly labeled data to
novel categories in the target domain that contains only few labeled data. The
knowledge transfer biases the novel category predictions and prevents the model
from overfitting. The knowledge is encoded using a neural-network-based prior
that transforms a data sample to its corresponding class prototype. This neural
network is trained from the source-domain data and applied to the target-domain
data, where it transforms the few-shot samples to the novel-class prototypes
for better recognition performance. The few-shot learning problem is then
extended to the situation, where we do not have access to the source-domain
data but only have access to the source-domain class prototypes. In this limited
information setting, parametric neural-network-based priors would overfit to
the source-class prototypes and hence we seek a non-parametric-based prior
using manifolds. A piecewise linear manifold is used as a structural prior to
fit the source-domain-class prototypes. This structure is extended to the
target domain, where the novel-class prototypes are found by projecting the
few-shot samples onto the manifold. Finally, the zero-shot learning problem is
addressed, which is an extreme case of the few-shot learning problem where we
do not have any labeled data in the target domain. However, we have high-level
information for both the source and target domain categories in the form of
semantic descriptors. We learn the relation between the sample space and the
semantic space, using a regularized neural network so that classification of
the novel categories can be carried out in a common representation space. This
same neural network is then used in the target domain to relate the two spaces.
In case we want to generate data for the novel categories in the target domain,
we can use a constrained generative adversarial network instead of a
traditional neural network. Thus, we use structural priors like graphs, neural
networks and manifolds to relate various data entities like samples, prototypes
and semantics for these different transfer learning sub-problems. We explore
additional post-processing steps like pseudo-labeling, domain adaptation and
calibration and enforce algorithmic and architectural constraints to further
improve recognition performance. Experimental results on standard transfer
learning image recognition datasets produced competitive results with respect
to previous work. Further experimentation and analyses of these methods
provided better understanding of machine learning as well.

Machine learning is widely used to extract meaningful information from video, images, audio, text, and other multimedia data.  Through a hierarchical structure, modern neural networks coupled with backpropagation learn to extract information from large amounts of data and to perform specific tasks such as classification or regression. In this thesis, we explore various approaches to multimedia analytics with neural networks. We present several image synthesis and rendering techniques to generate new images for training neural networks. Furthermore, we present multiple neural network architectures and systems for commercial logo detection, 3D pose estimation and tracking, deepfakes detection, and manipulation detection in satellite images.

This dissertation proposes an approach to reduce the cost of manual inspections for as large a number of false positive warnings that are being reported by Static Code Analysis (SCA) tools as much as possible using Machine Learning (ML) techniques. The proposed approach neither assume to use the particular SCA tools nor depends on the specific programming language used to write the target source code or the application. To reduce the number of false positive warnings we first evaluated a number of SCA tools in terms of software engineering metrics using a highlighted synthetic source code named the Juliet test suite. From this evaluation, we concluded that the SCA tools report plenty of false positive warnings that need a manual inspection. Then we generated a number of datasets from the source code that forced the SCA tool to generate either true positive, false positive, or false negative warnings. The datasets, then, were used to train four of ML classifiers in order to classify the collected warnings from the synthetic source code. From the experimental results of the ML classifiers, we observed that the classifier that built using the Random Forests

(RF) technique outperformed the rest of the classifiers. Lastly, using this classifier and an instance-based transfer learning technique, we ranked a number of warnings that were aggregated from various open-source software projects. The experimental results show that the proposed approach to reduce the cost of the manual inspection of the false positive warnings outperformed the random ranking algorithm and was highly correlated with the ranked list that the optimal ranking algorithm generated.

Phenylketonuria (PKU) is an inherited metabolic disorder affecting 1 in every 10,000 to 15,000 newborns in the United States every year. Caused by a genetic mutation, PKU results in an excessive build up of the amino acid Phenylalanine (Phe) in the body leading to symptoms including but not limited to intellectual disability, hyperactivity, psychiatric disorders and seizures. Most PKU patients must follow a strict diet limited in Phe. The aim of this research study is to formulate, implement and compare techniques for Phe estimation in commercial foods using the information on the food label (Nutritional Fact Label and ordered ingredient list). Ideally, the techniques should be both accurate and amenable to a user friendly implementation as a Phe calculator that would aid PKU patients monitor their dietary Phe intake.

The first approach to solve the above problem is a mathematical one that comprises three steps. The three steps were separately proposed as methods by Jieun Kim in her dissertation. It was assumed that the third method, which is more computationally expensive, was the most accurate one. However, by performing the three methods subsequently in three different steps and combining the results, we actually obtained better results than by merely using the third method.

The first step makes use of the protein content in the foods and Phe:protein multipliers. The second step enumerates all the ingredients in the food and uses the minimum and maximum Phe:protein multipliers of the ingredients along with the protein content. The third step lists the ingredients in decreasing order of their weights, which gives rise to inequality constraints. These constraints hold assuming that there is no loss in the preparation process. The inequality constraints are optimized numerically in two phases. The first involves nutrient content estimation by approximating the ingredient amounts. The second phase is a refinement of the above estimates using the Simplex algorithm. The final Phe range is obtained by performing an interval intersection of the results of the three steps. We implemented all three steps as web applications. Our proposed three-step method yields a high accuracy of Phe estimation (error <= +/- 13.04mg Phe per serving for 90% of foods).

The above mathematical procedure is contrasted against a machine learning approach that uses the data in an existing database as training data to infer the Phe in any given food. Specifically, we use the K-Nearest Neighbors (K-NN) classification method using a feature vector containing the (rounded) nutrient data. In other words, the Phe content of the test food is a weighted average of the Phe values of the neighbors closest to it using the nutrient values as attributes. A four-fold cross validation is carried out to determine the hyper-parameters and the training is performed using the United States Department of Agriculture (USDA) food nutrient database. Our tests indicate that this approach is not very accurate for general foods (error <= +/- 50mg Phe per 100g in about 38% of the foods tested). However, for low-protein foods which are typically consumed by PKU patients, the accuracy increases significantly (error <= +/- 50mg Phe per 100g in over 77% foods).

The machine learning approach is more user-friendly than the mathematical approach. It is convenient, fast and easy to use as it takes into account just the nutrient information. In contrast, the mathematical method additionally takes as input a detailed ingredient list, which is cumbersome to be located in a food database and entered as input. However, the Mathematical method has the added advantage of providing error bounds for the Phe estimate. It is also more accurate than the ML method. This may be due to the fact that for the ML method, the nutrition facts alone are not sufficient to estimate Phe and that additional information like the ingredients list is required.

This dissertation combines stakeholder and analytical intelligence for consensus decision-making via an interactive optimization process. This dissertation outlines techniques for developing user models of subjective criteria of human stakeholders for an environmental decision support system called WRESTORE. The dissertation compares several user modeling techniques and develops methods for incorporating such user models selectively for interactive optimization, combining multiple objective and subjective criteria.

This dissertation describes additional functionality for our watershed planning system, called WRESTORE (Watershed REstoration Using Spatio-Temporal Optimization of REsources) (http://wrestore.iupui.edu). Techniques for performing the interactive optimization process in the presence of limited data are described. This work adds a user modeling component that develops a computational model of a stakeholder’s preferences and then integrates the user model component into the decision support system.

Our system is one of many decision support systems and is dependent upon stake- holder interaction. The user modeling component within the system utilizes deep learning, which can be challenging with limited data. Our work integrates user models with limited data with application-specific techniques to address some of these challenges. The dissertation describes steps for implementing accurate virtual stakeholder models based on limited training data.

Another method for dealing with limited data, based upon computing training data uncertainty, is also presented in this dissertation. Results presented show more stable convergence in fewer iterations when using an uncertainty-based incremental sampling method than when using stability based sampling or random sampling. The technique is described in additional detail.

The dissertation also discusses non-stationary reinforcement-based feature selection for the interactive optimization component of our system. The presented results indicate that the proposed feature selection approach can effectively mitigate against superfluous and adversarial dimensions which if left untreated can lead to degradation in both computational performance and interactive optimization performance against analytically determined environmental fitness functions.

The contribution of this dissertation lays the foundation for developing a framework for multi-stakeholder consensus decision-making in the presence of limited data.