Browsing by Subject "Machine Learning"
Now showing 1 - 8 of 8
Results Per Page
Sort Options
Item Acoustic Based Sketch Recognition(2012-10-19) Li, WenzheSketch recognition is an active research field, with the goal to automatically recognize hand-drawn diagrams by a computer. The technology enables people to freely interact with digital devices like tablet PCs, Wacoms, and multi-touch screens. These devices are easy to use and have become very popular in market. However, they are still quite costly and need more time to be integrated into existing systems. For example, handwriting recognition systems, while gaining in accuracy and capability, still must rely on users using tablet-PCs to sketch on. As computers get smaller, and smart-phones become more common, our vision is to allow people to sketch using normal pencil and paper and to provide a simple microphone, such as one from their smart-phone, to interpret their writings. Since the only device we need is a single simple microphone, the scope of our work is not limited to common mobile devices, but also can be integrated into many other small devices, such as a ring. In this thesis, we thoroughly investigate this new area, which we call acoustic based sketch recognition, and evaluate the possibilities of using it as a new interaction technique. We focus specifically on building a recognition engine for acoustic sketch recognition. We first propose a dynamic time wrapping algorithm for recognizing isolated sketch sounds using MFCC(Mel-Frequency Cesptral Coefficients). After analyzing its performance limitations, we propose improved dynamic time wrapping algorithms which work on a hybrid basis, using both MFCC and four global features including skewness, kurtosis, curviness and peak location. The proposed approaches provide both robustness and decreased computational cost. Finally, we evaluate our algorithms using acoustic data collected by the participants using a device's built-in microphone. Using our improved algorithm we were able to achieve an accuracy of 90% for a 10 digit gesture set, 87% accuracy for the 26 English characters and over 95% accuracy for a set of seven commonly used gestures.Item Activity retrieval in closed captioned videos(2009-08) Gupta, Sonal; Mooney, Raymond J. (Raymond Joseph); Grauman, KristenRecognizing activities in real-world videos is a difficult problem exacerbated by background clutter, changes in camera angle & zoom, occlusion and rapid camera movements. Large corpora of labeled videos can be used to train automated activity recognition systems, but this requires expensive human labor and time. This thesis explores how closed captions that naturally accompany many videos can act as weak supervision that allows automatically collecting 'labeled' data for activity recognition. We show that such an approach can improve activity retrieval in soccer videos. Our system requires no manual labeling of video clips and needs minimal human supervision. We also present a novel caption classifier that uses additional linguistic information to determine whether a specific comment refers to an ongoing activity. We demonstrate that combining linguistic analysis and automatically trained activity recognizers can significantly improve the precision of video retrieval.Item Exploration, Registration, and Analysis of High-Throughput 3D Microscopy Data from the Knife-Edge Scanning Microscope(2014-04-25) Sung, ChulAdvances in high-throughput, high-volume microscopy techniques have enabled the acquisition of extremely detailed anatomical structures on human or animal organs. The Knife-Edge Scanning Microscope (KESM) is one of the first instruments to produce sub-micrometer resolution ( ~1 ?m^(3)) data from whole small animal brains. We successfully imaged, using the KESM, entire mouse brains stained with Golgi (neuronal morphology), India ink (vascular network), and Nissl (soma distribution). Our data sets fill the gap of most existing data sets which have only partial organ coverage or have orders of magnitude lower resolution. However, even though we have such unprecedented data sets, we still do not have a suitable informatics platform to visualize and quantitatively analyze the data sets. This dissertation is designed to address three key gaps: (1) due to the large volume (several tera voxels) and the multiscale nature, visualization alone is a huge challenge, let alone quantitative connectivity analysis; (2) the size of the uncompressed KESM data exceeds a few terabytes and to compare and combine with other data sets from different imaging modalities, the KESM data must be registered to a standard coordinate space; and (3) quantitative analysis that seeks to count every neuron in our massive, growing, and sparsely labeled data is a serious challenge. The goals of my dissertation are as follows: (1) develop an online neuro-informatics framework for efficient visualization and analysis of the multiscale KESM data sets, (2) develop a robust landmark-based 3D registration method for mapping the KESM Nissl-stained entire mouse data into the Waxholm Space (a canonical coordinate system for the mouse brain), and (3) develop a scalable, incremental learning algorithm for cell detection in high-resolution KESM Nissl data. For the web-based neuroinformatics framework, I prepared multi-scale data sets at different zoom levels from the original data sets. And then I extended Google Maps API to develop atlas features such as scale bars, panel browsing, and transparent overlay for 3D rendering. Next, I adapted the OpenLayers API, which is a free mapping and layering API supporting similar functionality as the Google Maps API. Furthermore, I prepared multi-scale data sets in vector-graphics to improve page loading time by reducing the file size. To better appreciate the full 3D morphology of the objects embedded in the data volumes, I developed a WebGL-based approach that complements the web-based framework for interactive viewing. For the registration work, I adapted and customized a stable 2D rigid deformation method to map our data sets to the Waxholm Space. For the analysis of neuronal distribution, I designed and implemented a scalable, effective quantitative analysis method using supervised learning. I utilized Principal Components Analysis (PCA) in a supervised manner and implemented the algorithm using MapReduce parallelization. I expect my frameworks to enable effective exploration and analysis of our KESM data sets. In addition, I expect my approaches to be broadly applicable to the analysis of other high-throughput medical imaging data.Item Investigations Into Using Machine Learning Models to Automate the Sorting of Digitized Texas State Publications(Texas Digital Library, 2023-05-16) Rikka, PraneethOver the past ten years the UNT Libraries has been digitizing Texas State Publications it receives from the Texas State Library and Archives Commission as part of the Texas State Depository program. During this time, over 19,000 items have been digitized and made available in The Portal to Texas History’s Texas State Publications Collection (https://texashistory.unt.edu/explore/collections/TXPUB/). Each year, batches of publications are sent to a digitization vendor, digitized, and sent back to UNT where each publication is sorted so that similar items are grouped together to assist in metadata creation. This sorting usually happens with sets of over 1,000 publications at a time. The manual sorting process is time consuming and requires expert knowledge of the subject matter. Recent advances in machine learning offer an automated approach to this manual sorting of documents. This poster presents a research project to build and test a classification model to assist librarians in the sorting of digitized Texas State Publications into groups. It discusses the labeled dataset that was created to test different machine learning approaches and presents the findings of text-based and image-based classification models. We hope that this poster encourages others in two specific ways, first to build datasets that highlight specific problems in the library and archives space that can be worked on by students interested in real world problems, and second, to think about processes that exist in their institution that might benefit from judicious use of machine learning to complement human decisions in making resources available for users.Item Machine Learning Based Classification of Textual Stimuli to Promote Ideation in Bioinspired Design(2013-08-09) Glier, Michael WBioinspired design uses biological systems to inspire engineering designs. One of bioinspired design?s challenges is identifying relevant information sources in biology for an engineering design task. Currently information can be retrieved by searching biology texts or journals using biology-focused keywords that map to engineering functions. However, this search technique can overwhelm designers with unusable results. This work explores the use of text classification tools to identify relevant biology passages for design. Further, this research examines the effects of using biology passages as stimuli during idea generation. Four human-subjects studies are examined in this work. Two surveys are performed in which participants evaluate sentences from a biology corpus and indicate whether each sentence prompts an idea for solving a specific design problem. The surveys are used to develop and evaluate text classification tools. Two idea generation studies are performed in which participants generate and record solutions for designing a corn shucker using either different sets of biology passages as design stimuli, or no stimuli. Based 286 sentences from the surveys, a k Nearest Neighbor classifier is developed that is able to identify helpful sentences relating to the function ?separate? with a precision of 0.62 and recall of 0.48. This classifier could potentially double the number of helpful results found using a keyword search. The developed classifier is specific to the function ?separate? and performs poorly when used for another function. Classifiers developed using all sentences and participant responses from the surveys are not able to reliably identify helpful sentences. From the idea generation studies, we determine that using any biology passages as design stimuli increases the quantity and variety of participant solutions. Solution quantity and variety are also significantly increased when biology passages are presented one at a time instead of all at once. Quality and variety are not significantly affected by the presence of design stimuli. Biological stimuli are also found to lead designers to types of solution that are not typically produced otherwise. This work develops a means for designers to find more useful information when searching biology and demonstrates several ways that biology passages can improve ideation.Item Metrics for sampling-based motion planning(2009-05-15) Morales Aguirre, Marco AntonioA motion planner finds a sequence of potential motions for a robot to transit from an initial to a goal state. To deal with the intractability of this problem, a class of methods known as sampling-based planners build approximate representations of potential motions through random sampling. This selective random exploration of the space has produced many remarkable results, including solving many previously unsolved problems. Sampling-based planners usually represent the motions as a graph (e.g., the Probabilistic Roadmap Methods or PRMs), or as a tree (e.g., the Rapidly exploring Random Tree or RRT). Although many sampling-based planners have been proposed, we do not know how to select among them because their different sampling biases make their performance depend on the features of the planning space. Moreover, since a single problem can contain regions with vastly different features, there may not exist a simple exploration strategy that will perform well in every region. Unfortunately, we lack quantitative tools to analyze problem features and planners performance that would enable us to match planners to problems. We introduce novel metrics for the analysis of problem features and planner performance at multiple levels: node level, global level, and region level. At the node level, we evaluate how new samples improve coverage and connectivity of the evolving model. At the global level, we evaluate how new samples improve the structure of the model. At the region level, we identify groups or regions that share similar features. This is a set of general metrics that can be applied in both graph-based and tree-based planners. We show several applications for these tools to compare planners, to decide whether to stop planning or to switch strategies, and to adjust sampling in different regions of the problem.Item Reinforcement Learning Control with Approximation of Time-Dependent Agent Dynamics(2013-04-30) Kirkpatrick, KentonReinforcement Learning has received a lot of attention over the years for systems ranging from static game playing to dynamic system control. Using Reinforcement Learning for control of dynamical systems provides the benefit of learning a control policy without needing a model of the dynamics. This opens the possibility of controlling systems for which the dynamics are unknown, but Reinforcement Learning methods like Q-learning do not explicitly account for time. In dynamical systems, time-dependent characteristics can have a significant effect on the control of the system, so it is necessary to account for system time dynamics while not having to rely on a predetermined model for the system. In this dissertation, algorithms are investigated for expanding the Q-learning algorithm to account for the learning of sampling rates and dynamics approximations. For determining a proper sampling rate, it is desired to find the largest sample time that still allows the learning agent to control the system to goal achievement. An algorithm called Sampled-Data Q-learning is introduced for determining both this sample time and the control policy associated with that sampling rate. Results show that the algorithm is capable of achieving a desired sampling rate that allows for system control while not sampling ?as fast as possible?. Determining an approximation of an agent?s dynamics can be beneficial for the control of hierarchical multiagent systems by allowing a high-level supervisor to use the dynamics approximations for task allocation decisions. To this end, algorithms are investigated for learning first- and second-order dynamics approximations. These algorithms are respectively called First-Order Dynamics Learning and Second-Order Dynamics Learning. The dynamics learning algorithms are evaluated on several examples that show their capability to learn accurate approximations of state dynamics. All of these algorithms are then evaluated on hierarchical multiagent systems for determining task allocation. The results show that the algorithms successfully determine appropriated sample times and accurate dynamics approximations for the agents investigated.Item Spammer Detection on Online Social Networks(2012-12-04) Amlesahwaram, Amit AnandTwitter with its rising popularity as a micro-blogging website has inevitably attracted attention of spammers. Spammers use myriad of techniques to lure victims into clicking malicious URLs. In this thesis, we present several novel features capable of distinguishing spam accounts from legitimate accounts in real-time. The features exploit the behavioral and content entropy, bait-techniques, community-orientation, and profile characteristics of spammers. We then use supervised learning algorithms to generate models using the proposed features and show that our tool, spAmbush, can detect spammers in real-time. Our analysis reveals detection of more than 90% of spammers with less than five tweets and more than half with only a single tweet. Our feature computation has low latency and resource requirement. Our results show a 96% detection rate with only 0.01% false positive rate. We further cluster the unknown spammers to identify and understand the prevalent spam campaigns on Twitter.