2016 Computer Science Abstracts

Poster Number: CSE-01

Title: A High Performance Block Eigensolver for Nuclear Configuration Interaction Calculations

Authors: Md Afibuzzaman; Hasan Metin Aktulga

Abstract: High accuracy prediction on the properties of light atomic nuclei using Configuration Interaction(CI) requires computing some extremal eigenpairs of the many-body Hamilton matrix, H. H is a sparse matrix, and although a Lanczos based eigensolver is commonly used for symmetric sparse eigenvalue problems, one can use the Locally Optimal Block Preconditioned Conjugate Gradient (LOBPCG) as it requires the multiplication of Sparse Matrix with Multiple vectors (SpMM). In SpMM, one can make use of the increased data locality and obtain much higher performance. A significant time is spent on multiplication of the H sparse matrix with multiple vectors and existing implementations of these multiplications do not perform as expected.  Here we analyze 4 different ways to implement SpMM and the transpose operation, SpMM_T. We base our implementation on the Compressed Sparse Blocks (CSB) format and target multicore architectures. The four implementations we consider here are: the traditional CSR (Compressed Sparse Row), Row partitioning algorithm, CSB using OpenMP and CSB using Cilk runtime environment. We also develop and analyze a performance model that allows us estimate the performance comparisons of our implementations. Extensive performance analysis show that the new CSB/OpenMP implementation achieves 3-4x speedup for SpMM and SpMM_T operations over good implementations based on CSR. Using a block eigensolver with optimized SpMM operations, we can attain 1.5-2x speed up in the overall execution time of the Lanczos based eigensolver used in MFDn.


Poster Number: CSE-02

Title: Intelligent and Automatic Quantification of in vivo Cells in MRI

Authors: Muhammad Afridi; Steven Hoffman; Arun Ross; Xiaoming Liu; Erik Shapiro

Abstract: Cell-based therapy (CBT) is emerging as a promising solution for a large number of serious health issues including organ transplant, cancer and brain injuries with limited success in humans. The key hurdle is our inability to determine the number and location of cells transplanted in live organs (in vivo) which severely affects our understanding about in vivo cell behavior. Therefore, experts conduct an MRI of the affected organ and then manually locate the transplanted cells to understand their behavior. However, manual enumeration in 3D MRI is a tedious task that is prone to subjectivity and inaccuracy. Hence, performing an essential large scale MRI analysis to understand in vivo cell behavior is clearly infeasible which directly hinders the success of CBT in humans. This study presents the first comprehensive research on how a computer based accurate, automatic and intelligent cell quantification approach can be developed for MRI scans. The proposed approach utilizes a deep learning based convolutional neural network (CNN) framework for accurately learning cell features in MRI and further exploits the principles of transfer learning via CNNs to learn accurately with only small amount of available medical data. Additionally, this study also shows how CNNs can incorporate learning from human labeling behavior to learn more accurate cell detection models. Comprehensive experimental evaluation using more than 100,000 testing samples show the proposed approach performs with an accuracy of up to 99.8% in vitro and 97.3% in vivo which was significantly superior to the traditional computer-vision and other CNN based approaches.

This work was supported in part by We are thankful to the support provided by NIH grants DP2 OD004362 (EMS), R21 CA185163 (EMS), R01 DK107697 (EMS), R21 AG041266 (EMS).


Poster Number: CSE-03

Title: A Machine-to-Machine Outsourcing Approach to Monitoring Quality of Experience for Operational Cellular Networks

Authors: Faraz Ahmed; Jeffrey Erman; Zihui Ge; Alex X. Liu; Jia Wang; He Yan

Abstract: Cellular data networks are being increasingly used to connect to the Internet. to provide seamless Internet access with high quality of experience, it is crucial for cellular network operators to monitor network health and assess service quality perceived by its customers. Machine-to-Machine (M2M) devices communicate using the same cellular network as human operated devices such as mobile phones. This provides an unprecedented opportunity for cellular network operators to use M2M devices as free sensors on the field to measure what end-users experience. In this paper, we propose to out source network service quality monitoring to M2M devices. This approach provides a brand-new view that is much closer to the end-users' experience. We present our design and prototype of a monitoring system called M2MScan for a large-scale operational cellular network. We first identify a candidate set of M2M devices with highly predictable mobility and communication patterns using data collected from a large-scale cellular provider in North America. We then use these M2M devices to detect cell tower outage and estimate customer impact of cell tower outages. We evaluate M2MScan by using cell tower outages in a 2-months time window and our operational experience reveals that: (i) customer's experience can be measured at a fine-granularity by monitoring M2M device communication; (ii) customer experience during network outages varies both in space and time; and (iii) customer experience estimates can be used by cellular network provider to improve network coverage and network resiliency. to the best of our knowledge, this is the first work that employs M2M devices to monitor operational cellular networks at a large scale.


Poster Number: CSE-04

Title: Epileptic Seizure Inference

Authors: Atra Akandeh; Fathi Salem

Abstract: Epilepsy is one of the most debilitating neurological disorders affecting about 70 million people in the world (according to the International League Against Epilepsy). Epilepsy entails a temporal change in the brain’s electrical activity that expresses itself in motor, psychic, and sensory manifestations associated with spasms. Diagnosis of epilepsy is usually performed by analyzing electroencephalogram (EEG) signals, as well as the patient’s neurological behavior. However, this approach requires time-consuming recordings and their analysis by an expert. Automated analysis of EEG recordings to assist in the diagnosis and inference of epilepsy has continued to this day. The most common methods for seizure detection are based on principal and independent component analysis, clustering techniques, and data mining classification techniques, and recently feedforward neural networks. In this project, we propose an automatic analysis system based on deep neural networks in order to infer and classify the brain firing signals into one of four possible epileptic states: inter-ictal, pre-ictal, ictal and post-ical. This approach will integrate two main components: feature extraction and inference/classification. The methodology will also incorporate non-convex optimization methods using homotopic diffusion equations to improve the deep network performance.


Poster Number: CSE-05

Title: Keystrokes Recognition Using Wi-Fi Signals

Authors: Kamran Ali; Alex X. Liu; Wei Wang; Muhammad Shahzad

Abstract: Keystroke privacy is critical for ensuring the security of computer systems and the privacy of human users as what being typed could be passwords or privacy sensitive information. In this paper, we show for the first time that WiFi signals can also be exploited to recognize keystrokes. The intuition is that while typing a certain key, the hands and fingers of a user move in a unique formation and direction and thus generate a unique pattern in the time-series of Channel State Information (CSI) values, which we call CSI-waveform for that key. In this paper, we propose a WiFi signal based keystroke recognition system called WiKey. WiKey consists of two Commercial Off-The-Shelf (COTS) WiFi devices, a sender (such as a router) and a receiver (such as a laptop). The sender continuously emits signals and the receiver continuously receives signals. When a human subject types on a keyboard, WiKey recognizes the typed keys based on how the CSI values at the WiFi signal receiver end. We implemented the WiKey system using a TP-Link WiFi router and a Lenovo X200 laptop. WiKey achieves more than 97.5% detection rate for detecting the keystroke and 96.4% recognition accuracy for classifying single keys. In real-world experiments, WiKey can recognize keystrokes in a continuously typed sentence with an accuracy of 93.5% for reasonable typing speeds.


Poster Number: CSE-06

Title: WiGesture: Facial Gesture Recognition Using WiFi Signals

Authors: Salman Ali; Kamran Ali; Alex X. Liu; Wei Wang

Abstract: Recently WiFi signals have been used for detecting human activities based on disturbances created in Received Signal Strength and Channel Side Information (CSI). Variations in CSI values for Multiple Input Multiple Output (MIMO) streams of WiFi devices with multiple antennas can be leveraged to detect and recognize various human or hon-human activities within the vicinity. In this paper, we propose WiGesture, a CSI based facial gesture recognition system that quantifies small variations in CSI caused by gestures related to pronunciation of common interjections. Detection of such interjections associated with a particular emotion or sentiment are particularly difficult to detect since the variations are mostly embedded as noise in the CSI variations. To achieve this, WiGesture needs to detect and analyze fine-grained signal reflections from multi-paths generated from facial movements. We not only propose a novel approach to differentiate gestures, we also present our work in a real case application of smartphone usage scenario. The implementation of WiGesture is done in real time on a commodity WiFi device with detection accuracy in the range of 80% using only 6 training samples.


Poster Number: CSE-08

Title: Automated Detection and Tracking of a Trailers Coupler

Authors: Yousef Atoum; Joseph Roth; Xiaoming Liu

Abstract: Our work aims at developing an effective and efficient computer vision (CV) system to continuously detect and track a trailers tongue in real time, for the purpose of simplifying the action of coupling the tongue of the trailer to a truck. Even for experienced truck drivers, attaching the trailer to the truck might be a challenging task. More recently, some newly manufactured cars contain a rear-view built-in camera that would help in this process. We propose to use the camera sensor in developing an automatic CV system to detect and track the trailers coupler. We explore a large variety of possible detectors and trackers including Correlation Filter (CF) and Convolutional Neural Networks (CNN). to evaluate our system, we collect data from several trailer dealer sites. The dataset contains 406 videos varying in trailer types (models, colors, and shapes), weather conditions (Sunny and cloudy), and different environments (streets, gravel, snow and dirt). The videos also have several challenges such as pose, scale and illumination variations. Extensive experimental results demonstrate the accuracy, robustness, and efficiency of our automated system.

This work was supported in part by General Motors (GM)


Poster Number: CSE-07

Title: Livestock Detection and Tracking in the Thermal Spectrum

Authors: Aaron Gonzales; Arun Ross

Abstract: Thermal cameras are becoming increasingly cost-effective, thereby making them more affordable in a number of applications. In this project, we will utilize thermal cameras as a surveillance instrument to monitor a livestock pen. In particular, we will develop automated computer vision and pattern recognition methods to detect and track livestock as well as intruders - both animal and human - attempting to compromise the security of the livestock pen. While extensive research on object detection and tracking has been conducted in the visible spectra, little work has been conducted in the thermal spectra. In this work, we will design a classifier that detects individual objects (animals/human) in an enclosure and labels them using a multi-class classifier. Next, we will develop a method to track individual objects in order to determine their intent (e.g., do they pose a threat to the other animals). Finally, we will acquire data using a thermal camera and evaluate the performance of the proposed methods. The proposed methods are expected to be resilient to a number of confounding factors such as occlusion, inclement weather, low-intensity lighting, etc.


Poster Number: CSE-09

Title: Automated Online Exam Proctoring

Authors: Yousef Atoum; Liping Chen; Alex Liu; Stephen Hsu; Xiaoming Liu

Abstract: Massive open online courses (MOOCs) and other forms of remote education continue to increase in popularity and reach. The ability to efficiently proctor remote online examinations is an important limiting factor to the scalability of this next stage in education. Presently, human proctoring is the most common approach of evaluation, by either requiring the test taker to visit an examination center, or by monitoring them visually and acoustically during exams via a webcam. However, such methods are labor-intensive and costly. In this work, we present a multimedia analytics system that performs automatic online exam proctoring. The system hardware includes one webcam, one wearcam, and a microphone, for the purpose of monitoring the visual and acoustic environment of the testing location. The system includes six basic components that continuously estimate the key behavior cues: user verification, text detection, voice detection, active window detection, gaze estimation and phone detection. By combining the continuous estimation components, and applying a temporal sliding window, we design higher-level features to classify whether the test taker is cheating at any moment during the exam. to evaluate our proposed system, we collect multimedia (audio and visual) data from 24 subjects performing various types of cheating while taking online exams. Extensive experimental results demonstrate the accuracy, robustness, and efficiency of our online exam proctoring system.

This work was supported in part by Michigan State University Targeted Support Grants for Technology Development (TSGTD) program.


Poster Number: CSE-10

Title: From Which Camera Did This Iris Image Come From?

Authors: Sudipta Banerjee; Arun Ross

Abstract: Iris recognition systems use images of the iris (which is the textured annular portion of the human eye) for human recognition. The proposed work attempts to automatically determine the identity of the sensor, i.e., the device, based on the raw image alone. Device identification is useful in applications where digital tampering is prevalent. Further, it can be used to invoke a specific set of image processing routines based on knowledge of the sensor that was used to acquire the image. Current research exploits the notion of Photo Response Non-Uniformity (PRNU) noise -  a distinctive sensor pattern embedded in the image - for device identification. However, the use of PRNU noise pattern has certain limitations. For example, the noise pattern can be contaminated by the scene details, thereby, impacting the device identification accuracy. In our work, we use the recently proposed Enhanced Sensor Pattern Noise (ESPN) to enhance the sensor pattern noise and subdue the scene content in the context of iris images acquired in the near-infrared spectrum. Experiments involving images from multiple iris sensors confirm the benefits of the proposed approach.


Poster Number: CSE-11

Title: Stochastic Convex Sparse Principle Component Analysis

Authors: Inci M. Baytas; Kaixiang Lin; Fei Wang; Jiayu Zhou; Anil K. Jain

Abstract: Principal Component Analysis (PCA) is a dimensionality reduction and data analysis tool commonly used in many areas. The main idea of PCA is to represent high dimensional data with a few representative components that capture most of the variance present in the data. However, there is an obvious disadvantage of traditional PCA when it is applied to analyze data where interpretability is important. In applications, where the features have some physical meaning, we lose the ability to interpret the principal components extracted by conventional PCA because each principal component is a linear combination of all the original features. For this reason, sparse PCA has been proposed to improve the interpretability of traditional PCA, by introducing sparsity to the loading vectors of principle components. In this paper, we propose a convex sparse principle component analysis (Cvx-SPCA), which leverages a proximal variance reduced stochastic scheme to achieve a geometric convergence rate. We further show that the convergence analysis can be significantly simplified by using a weak condition, allowing a broader class of objectives to be applied. The efficiency and effectiveness of the proposed method is demonstrated on a large scale electronic medical record cohort.


Poster Number: CSE-13

Title: Exploring Sex Prediction from a Near Infrared Iris Image

Authors: Denton Bobeldyk; Arun Ross

Abstract: Recent research has explored the possibility of automatically deducing the sex of an individual based on near infrared (NIR) images of the iris. This has benefits in the context of an iris biometric system, where an individual is recognized based on an NIR image of the iris. Most operational iris biometric systems typically acquire an image of the extended ocular region (rather than just the iris only) for processing. In this work, we investigate the sex-predictive  accuracy associated with three different regions: (a) the entire ocular region; (b) the iris-excluded ocular region; and (c) the iris-only region. We employ the BSIF texture operator (Binarized Statistical Image Feature) to extract features from these regions, and use a Support Vector Machine (SVM) to classify the extracted feature set as Male or Female. Experiments on a dataset containing 3314 images suggests that the iris region only provides modest sex-specific cues compared to the surrounding ocular region. This research underscores the importance of using the periocular region in iris recognition systems.


Poster Number: CSE-14

Title: FingerprintMash: Latent Fingerprint Value Determination by Expert Crowdsourcing

Authors: Tarang Chugh; Kai Cao; Anil K. Jain

Abstract: Automatic fingerprint identification systems (AFIS), first introduced in the early 1980s, are now used by virtually every law enforcement and forensic agency to identify victims and suspects. However, identifying a person, whether by latent experts or AFIS, based on latent (partial) fingerprints left at crime scenes continues to pose a challenge. This challenge is primarily due to the generally low quality or value of latent prints. Latent fingerprint value can be defined in terms of quality and quantity of information content (e.g., ridge clarity and no. of minutiae). In forensics, latents are typically categorized by experts based on their value: (i) value for identification (VID), (ii) value for exclusion only (VEO), and (iii) no value (NV). But, this manual process of value determination is subjective. In order to understand this subjectivity and to develop a baseline for evaluating automatic methods of value determination, our research has made the following contributions: (i) Designed a crowdsourcing tool, called FingerprintMash, which allows latent experts to assign an absolute and a relative value to a latent and a pair of latents, respectively, (ii) asked a pool of 33 fingerprint experts to assign values to a set of 100 latent pairs chosen randomly from our database of 516 latents, (iii) used matrix completion to infer latent rankings based on absolute and relative values, and (iv) utilized Multidimensional Scaling (MDS) to visualize the 516 x 516 similarity matrix of the 516 latents to determine the underlying factors that latent experts use to assign a value. Our analysis shows that (i) there is large variability in the values assigned by the expert crowd, (ii) crowdsourced value (median) performs better than value by a single examiner, in terms of predicting the AFIS performance, and (iii) Multidimensional Scaling enables us to understand the basis of latent experts’ value determination.


Poster Number: CSE-15

Title: Exploring Spatial Skills of Introductory Programming Students

Authors: Sarah Coburn; Mark Urban-Lurain

Abstract: Introductory programming courses are often perceived as difficult, frequently discouraging potentially talented students from continuing in the major. One potential reason for this difficulty is extraneous cognitive load: information that competes for mental resources with the essential concepts in introductory programming courses. We examine spatial visualization skills: the ability to perceive and mentally manipulate spatial objects. We propose that spatial visualization skills positively impact the ability to process visual information in working memory, therefore minimizing extraneous load and allowing mental resources to be dedicated to learning essential computer programming concepts. We present initial findings of student spatial skills before and after a semester in an introductory programming course, and the relationship with course performance. We discuss ongoing research on training student spatial skills, as well as future research for improving introductory programming curriculum by presenting concepts spatially, utilizing multiple channels of working memory.


Poster Number: CSE-16

Title: The Forge: Building Efficient Packet Classifiers

Authors: James Daly; Eric Torng

Abstract: Packet classification is a major component of network devices, such as firewalls and forwarding tables. Because these devices have real-time constraints, it is important that they are able to classify packets efficiently. If they do not, the entire network may become congested. For newer software-defined networks, fast updates are also important. We present a new classifier that provides both fast classification times and fast updates.


Poster Number: CSE-17

Title: A Bayesian Belief Fusion Framework for Integrating Match Scores with Auxiliary Information in Fingerprint Verification Systems

Authors: Yaohui Ding; Ajita Rattani; Arun Ross

Abstract: Recent research has addressed the robustness of fingerprint verification systems against spoof attacks by combining match scores with both liveness measures as well as image quality in a learning-based fusion framework. Designing such a fusion framework is challenging because quality and liveness measures can impact the match score and, therefore, the influence of these variables on the match score has to be modelled. We advance the state-of-the-art on this topic by proposing two Bayesian Belief Network (BBN) models that can utilize these measures effectively and appropriately model the relationship between quality, liveness measure and match scores. We show that the proposed BBN models result in consistently better matching performance than existing fusion frameworks.


Poster Number: CSE-18

Title: The Effects of Evolution and Spatial Structure on Diversity in Biological Reserves

Authors: Emily Dolson; Michael Wiser; Charles Ofria

Abstract: Conservation ecologists have long argued over the best way of placing reserves across an environment to maximize population diversity. Many have studied the effect of protecting many small regions of an ecosystem vs. a single large region, with varied results.However, this research tends to ignore evolutionary dynamics under the rationale that the spatiotemporal scale required is prohibitive. We used the Avida digital evolution research platform to overcome this barrier and study the response of phenotypic diversity to eight different reserve placement configurations. The capacity for mutation, and therefore evolution, substantially altered the dynamics of diversity in the population. When mutations were allowed, reserve configurations involving a greater number of consequently smaller reserves were substantially more effective at maintaining existing diversity and generating new diversity. However, when mutations were disallowed, reserve configuration had little effect on diversity generation and maintenance. While further research is necessary before translating these results into policy decisions, this study demonstrates the importance of considering evolution when making such decisions and suggests that a larger number of smaller reserves may have evolutionary benefits.

This work was supported in part by - BEACON Center - NSF Doctoral Fellowship


Poster Number: CSE-19

Title: Causality of Verbs for Grounded Language Understanding

Authors: Qiaozi Gao; Malcolm Doering; Shaohua Yang; Joyce Chai

Abstract: Linguistics studies have shown that concrete action verbs often denote some Change of State (CoS) as the result of an action. However, the causality of verbs and its potential connection with the physical world has not been systematically explored. to address this limitation, this work presents our study on verb causality modeling. We first conducted a crowd-sourcing study to identify potential categories of causality for a selected set of verbs. Associated with these categories, we then defined a set of rule-based detectors using visual perception information. Our empirical results have demonstrated that these simple detectors can be directly applied for grounding language to perception (i.e., grounding semantic roles of verbs to perceived objects) and achieve competitive performance. When the training data is available, the association between the detectors and the verbs can be further learned, which is shown to achieve significant performance gain compared to a state-of-the-art approach in grounding.

This work was supported in part by NSF and DARPA.


Poster Number: CSE-20

Title: Mapping the Genomic Architecture of Adaptive Traits with Interspecific Introgressive Origin

Authors: Hussein Hejase; Kevin Liu

Abstract: Introgression involves the transfer of genetic information from one species/population to another as a result of hybridization and repeated backcrossing. Introgression has played a key role in the evolution of novel traits in many different organisms, including adaptation to high-altitude environments in humans, evolution of mimetic butterfly wing patterns, and pesticide resistance in house mice. The goal of this work is to identify the genomic architecture of introgressed traits. We use association mapping, which pinpoints statistical associations between genotypic and trait characters, to uncover the underlying genetic factors contributing to variation in a trait of interest. One of the issues that need to be addressed when conducting an association mapping study is sample relatedness, which induces spurious associations between genotypic and trait characters when the evolutionary relatedness among samples is not accounted for or wrong. To address this issue, we introduce Coal-Map 2, a new method that combines a linear mixed model to evaluate the relationship between genotypic and trait characters with an evolutionary model to capture complex sample relatedness. We explore the performance of Coal-Map 2 using an extensive performance study. We find that Coal-Map 2 significantly outperforms state-of-the-art methods (including EIGENSTRAT and our previously introduced method Coal-Map) both in terms of true positive rate and false positive rate. At a typical false positive rate of 5%, Coal-Map 2's true positive rate was better than EIGENSTRAT and Coal-Map by 20% and 10%, respectively.

This work was supported in part by This work was partially supported by Grant CCF-1565719 from the National Science Foundation and by startup funds from Michigan State University (to K.L.).


Poster Number: CSE-21

Title: Towards a Truthful Online Spectrum Auction with Dynamic Demand and Supply

Authors: Chowdhury Hyder; Thomas Jeitschko; Li Xiao

Abstract: In spectrum trading, secondary users bid for the spectrum units being made available by the primary users. Auction theory has been widely applied to improve spectrum allocation in such spectrum trading scenarios. However in contrast to reality, most of the research work assume either static user population or static spectrum supply or both. In this work, we investigate a realistic dynamic auction environment where secondary users with diverse delay bounds arrive dynamically and spectrum becomes available at random. We propose a priority ranking based online auction mechanism that prevents bidders from gaining advantage by misreporting information. We prove that the proposed auction mechanism is truthful and individual rational. We illustrate the properties of the mechanism in terms of spectrum utilization rate, bidder satisfaction rate, and average bidder utility through extensive simulations.

This work was supported in part by NSF


Poster Number: CSE-22

Title: Large-Pose Face Alignment via CNN-Based Dense 3D Model Fitting

Authors: Amin Jourabloo; Xiaoming Liu

Abstract: Face alignment aims to estimate the locations of a set of landmarks for a given image. This problem has received much attention as evidenced by the recent advancement in both the methodology and performance. However, most of the existing works neither explicitly handle face images with arbitrary poses, nor perform large-scale experiments on non-frontal and profile face images. In order to address these limitations, we proposed a novel face alignment algorithm that estimates both 2D and 3D landmarks and their 2D visibilities for a face image with an arbitrary pose. By integrating a 3D point distribution model, a cascaded coupled-regressor approach is designed to estimate both the camera projection matrix and the 3D landmarks. Furthermore, the 3D model also allows us to automatically estimate the 2D landmark visibilities via surface normal. We use a substantially larger collection of all-pose face images to evaluate our algorithm and demonstrate superior performances than the state-of-the-art methods.


Poster Number: CSE-23

Title: The Evolutionary Origins of Phenotypic Plasticity

Authors: Alexander Lalejini; Charles Ofria

Abstract: Many effective and innovative survival mechanisms used by natural organisms rely on the capacity for phenotypic plasticity; that is, the ability of a genotype to alter how it is expressed based on the current environmental conditions. Understanding the evolution of phenotypic plasticity is an important step towards understanding the origins of many types of biological complexity, as well as to meeting challenges in evolutionary computation where dynamic solutions are required. In this work, we leverage the Avida Digital Evolution Platform to experimentally explore the selective pressures and evolutionary pathways that lead to phenotypic plasticity. We present evolved lineages wherein unconditional traits tend to evolve first; next, imprecise forms of phenotypic plasticity often appear before optimal forms finally evolve. We visualize the phenotypic states traversed by evolved lineages across environments with differing rates of mutations and environmental change. We see that under all conditions, populations can fail to evolve phenotypic plasticity, instead relying on mutation-based solutions.

This work was supported in part by This work was supported in part the US National Science Foundation under cooperative agreement No. DBI-0939454 and by Michigan State University through a fellowship for Lalejini.


Poster Number: CSE-24

Title: Multi-Task Feature Interaction Learning

Authors: Kaixiang Lin; Jianpeng Xu; Shuiwang Ji; Jiayu Zhou

Abstract: Linear models are widely used in various data mining and machine learning algorithms. One major limitation of such models is the lack of capability to capture predictive information from interactions between features. While introducing high-order feature interaction terms can overcome this limitation, this approach dramatically increases the model complexity and imposes significant challenges in the learning against overfitting. When there are multiple related learning tasks, feature interactions from these tasks are usually related and modeling such relatedness is the key to improve their generalization. In this paper, we propose a novel Multi-Task feature Interaction Learning (MTIL) framework to exploit the task relatedness from high-order feature interactions. Specifically, we collectively represent the feature interactions from multiple tasks as a tensor, and prior knowledge of task relatedness can be incorporated into different structured regularizations on this tensor. We formulate two concrete approaches under this framework, namely the shared interaction approach and the embedded interaction approach. The former assumes tasks share the same set of interactions, and the latter assumes feature interactions from multiple tasks share a common subspace. We have provided efficient algorithms for solving the two formulations. Extensive empirical studies on both synthetic and real datasets have demonstrated the effectiveness of the proposed framework.

This work was supported in part by Office of Naval Research and National Science Foundation.


Poster Number: CSE-25

Title: Inter-Femtocell Interference Identification and Resource Management

Authors: Chin-Jung Liu; Pei Huang; Li Xiao

Abstract: OFDMA femtocell is a promising technology to improve indoor wireless cellular network coverage in a cost-effective way. Large-scale deployment of femtocells in urban area is expected in the near future. However, inter-femtocell interference significantly limits the achievable throughput of a femtocell network. A typical approach to mitigate inter-cell interference is known as resource isolation, which aims at assigning non-overlapping resources to interfering femtocells. A major challenge for interference mitigation in femtocell networks is that the femtocells are often installed by end-consumers without any pre-planning. Very limited information about the femtocells is available, making it hard to decipher the inter-femtocell interference. Previous studies either take time to resolve collisions online or adopt a conservative approach to identify interferers. Although the latter approach avoids wasting time on resolving collisions, it may result in resource underutilization. In this paper, we propose an efficient method to identify inter-femtocell interference by analyzing the received patterns observed by mobile stations. We conducted experiments on GNU Radio/USRP to demonstrate that the proposed interference identification method can successfully identify real interferers while excluding non-interfering femto- cells from suspicious interfering femtocells. With the proposed interference identification, the resource allocation to the femtocells can achieve better fairness and higher throughput.


Poster Number: CSE-26

Title: Model Repair

Authors: Mohammad Roohitavaf; Sandeep Kulkarni

Abstract: Model repair tries to find an optimal balance between model checking -- which focuses on verifying the given model and model synthesis -- which focuses on designing a model that is correct by construction. In particular, model repair begins with an existing model that satisfies a subset of desired properties. Then it revises that model so that it preserves those properties while also satisfying some new properties such as fault-tolerance, stabilization, safety and liveness.


Poster Number: CSE-27

Title: Network Aware Task Scheduling in Data Centers

Authors: Ali Munir; Alex Liu

Abstract: Datacenters are being used as a critical infrastructure for high-revenue online services such as web search, social networking, and recommendation systems. For provisioning such large-scale online applications, datacenters face extreme challenges in providing desired user experience. These datacenter applications have very demanding latency requirements and even a small fraction of a second can make a quantifiable difference in user experience thus impacting the revenue. For example, Google observed a 20% traffic reduction from an extra 500ms, and Amazon found that every additional 100ms of latency costs them a 1% loss in business revenue. to improve the performance of these applications, existing datacenter schedulers optimize either the placement of tasks or the scheduling of network flows. The task scheduler strives to place tasks close to their input data to minimize network traffic, while assuming fair sharing of the network. The network scheduler strives to finish flows as quickly as possible based on their sources and destinations determined by the task scheduler. Inconsistent assumptions of the two schedulers can compromise the overall application performance. In this work, we propose NEAT, a task scheduling framework that leverages information from the underlying network scheduler to make task placement decisions. The core of NEAT is a task completion time predictor that estimates the completion time of a task under a given network condition and a given network scheduling policy. NEAT improves application performance by up to 4x for suboptimal network scheduling policies and up to 30% for optimal network scheduling policies.


Poster Number: CSE-28

Title: Toward Efficient Methods for Charge Equilibration in Polarizable, Reactive Molecular Dynamics Applications

Authors: Kurt A. O'Hearn; H. Metin Aktulga

Abstract: Polarizable, reactive methods, which incorporate range-limited quantum mechanics-like (QM) interactions based on bond-order potentials, have been shown to be an effective combination of the advantageous characteristics of QM and classical methods. The objective of this work is to enhance the efficiency of reactive force field molecular dynamics applications. Specifically, efforts are focused on optimization of the most costly phase of the PuReMD simulation, charge equilibration (QEq). Krylov subspace iterative approaches are employed for solving the large sparse symmetric linear systems defining the underlying QEq problem. The results of work on accelerating the convergence rate of these iterative techniques via incomplete LU (ILU) preconditioning techniques are presented. Ongoing investigations on parallelization of computation and application of preconditioning factors and the overarching solver are also discussed.


Poster Number: CSE-29

Title: Clustering Millions of Faces By Identity

Authors: Charles Otto; Dayong Wang; Anil Jain

Abstract: In this work, we attempt to address the following problem: Given a large number of unlabeled face images, cluster them into the individual identities present in this data. We consider this a relevant problem in different application scenarios ranging from social media to law enforcement. In large-scale scenarios the number of faces in the collection can be of the order of hundreds of million, while the number of clusters can range from a few thousand to millions–leading to difficulties in terms of both run-time complexity and evaluating clustering and per-cluster quality. A modified Rank-Order clustering algorithm is developed based on an approximate k-nearest-neighbor algorithm to achieve the desired scalabilityIn this work, we attempt to address the following problem: Given a large number of unlabeled face images, cluster them into the individual identities present in this data. We consider this a relevant problem in different application scenarios ranging from social media to law enforcement. In large-scale scenarios the number of faces in the collection can be of the order of hundreds of million, while the number of clusters can range from a few thousand to millions–leading to difficulties in terms of both run-time complexity and evaluating clustering and per-cluster quality. A modified Rank-Order clustering algorithm is developed based on an approximate k-nearest-neighbor algorithm to achieve the desired scalability. We cluster up to 123 million face images into over 10 million clusters, and analyze the results in terms of both external cluster quality measures (known face labels) and internal cluster quality measures (unknown face labels) and run-time. In terms of external cluster quality, our algorithm achieves an F-measure of 0.87 (range is [0,1]) on a small face dataset (LFW, consisting of 13K faces), and 0.27 on the largest dataset considered (123M faces). Additionally, we present preliminary work on video frame clustering (achieving 0.71 F-measure when clustering all frames in the YouTube Faces dataset). An internal per-cluster quality measure is developed which can be used to rank individual clusters and to automatically identify a subset of good quality clusters for manual investigation. We cluster up to 123 million face images into over 10 million clusters, and analyze the results in terms of both external cluster quality measures (known face labels) and internal cluster quality measures (unknown face labels) and run-time. In terms of external cluster quality, our algorithm achieves an F-measure of 0.87 (range is [0,1]) on a small face dataset (LFW, consisting of 13K faces), and 0.27 on the largest dataset considered (123M faces). Additionally, we present preliminary work on video frame clustering (achieving 0.71 F-measure when clustering all frames in the YouTube Faces dataset). An internal per-cluster quality measure is developed which can be used to rank individual clusters and to automatically identify a subset of good quality clusters for manual investigation.


Poster Number: CSE-30

Title: Secure Face Unlock: Robust Spoof Face Detection on Smartphones?

Authors: Keyurkumar Patel; Hu Han; Anil K. Jain

Abstract: With the wide deployment of face recognition systems in applications from de-duplication to mobile device unlocking, security against face spoofing attacks requires increased attention; such attacks can be launched via printed photos, video replays and 3D masks of a face. We address the problem of facial spoof detection against print (photo) and replay (photo or video) attacks based on the analysis of image aliasing (e.g., surface reflection, moire pattern, color distortion, and shape deformation) in spoof face images. Our application of interest is smartphone unlock, given that growing number of phones have face unlock and mobile payment capabilities. We develop an efficient face spoof detection system on an Android smartphone. The system was trained using an in-house database of real and spoof faces of 1,000 persons that included both print and replay attacks. Experimental results on public- domain face spoof databases, and the MSU USSA database show that the proposed approach is effective in face spoof detection for both cross-database and intra-database testing scenarios. User studies of our Android face spoof detection system involving 20 participants show that the proposed approach works very well in real application scenarios.


Poster Number: CSE-31

Title: iFrame: Dynamic Indoor Map Construction through Automatic Mobile Sensing

Authors: Chen Qiu; Matt Mutka

Abstract: Many pervasive applications depend upon maps for navigation and support of location based services. Maps are commonly available for outdoor pervasive applications from a variety of sources. An individual can determine their location outdoors on these maps via GPS. Indoor mobile applications may also need to know the layout of buildings, however indoor maps of buildings are less prevalent. Moreover, indoor maps may need to be dynamic and updated regularly because of the layout changes by people. We present iFrame, a dynamic approach that leverages existing mobile sensing capabilities for constructing indoor maps. We explore how iFrame users may collaborate and contribute to constructing 2 -dimensional indoor maps by merely carrying mobile devices. The iFrame approach consists of four steps: 1) Abstract the unknown indoor map as a matrix; 2) Leverage collaborating mobile devices that incorporate three mobile sensing technologies - accelerometers to support dead reckoning, Bluetooth RSSI detection, and WiFi RSSI detection; 3) Combine the three methods by Curve Fit Fusion, and 4) Extend iFrame from one room to a whole building by shadow rates and anchor points analysis. We conducted a deployment study that shows iFrame is a light-weight and unattended approach that provides a skeleton map of a real building effectively and automatically. The layouts of 12 rooms are reconstructed within 5-10 minutes. Changes of layout in indoor maps can be detected and the resolution of the reconstructed indoor floor plans can be improved when there is an increase in the number of cooperating users.

This work was supported in part by This work is supported in part by NSF Grant No. CNS-1320561.


Poster Number: CSE-32

Title: Network Completion with Provable Guarantees by Leveraging Side information

Authors: Abdol-Hossein Esfahanian; Dennis Ross; Farzan Masrour; Hayder Radha; Iman Barjasteh; Rana Forsati

Abstract: Link prediction is an important aspect of social network analysis and an area of key research within that is the network completion problem, where it is assumed that only a small sample of a is observed and we would like to infer the unobserved part of the network. In a typical network completion problem the standard methods, such as matrix completion, are inapplicable due the non-uniform sampling of observed links. This paper investigates the network completion problem and demonstrates that by effectively leveraging the side information about the nodes (such as the pairwise similarity), it is possible to predict the unobserved part of the network with high accuracy. to this end, we propose an efficient algorithm that decouples the completion from transduction stage to effectively exploit the similarity information. This crucial difference greatly boosts the performance where appropriate similarity information is used. The recovery error of the proposed algorithm is analyzed theoretically based on the richness of the similarity information and the size of the observed sub-network. to the best of our knowledge, this is the first algorithm that addresses the network completion with similarity of nodes with provable guarantees. Through extensive experiments on four real world datasets, we demonstrate that (i) leveraging side information in matrix completion by decoupling the completion from transduction significantly improves the link prediction performance, (ii) proposed two-stage method can deal with the cold-start problem that arises when a new entity enters the network, and (iii) our approach is scalable to large-scale networks.


Poster Number: CSE-33

Title: Adaptive 3D Face Reconstruction From Unconstrained Photo Collections

Authors: Joseph Roth; Yiying Tong; Xiaoming Liu

Abstract: This work presents a method for adaptive 3D face reconstruction from an unconstrained photo collection. Given a collection of “in-the-wild” face images captured under a variety of pose, expression, and illumination conditions, the algorithm produces a 3D face surface model of an individual along with albedo information. Motivated by the success of recent face reconstruction techniques on large photo collections, we extend prior work to adapt to low quality photo collections with fewer images. We achieve this by fitting a 3D Morphable Model to form a personalized template and developing a novel photometric stereo formulation, under a coarse to fine scheme. Superior experimental results are reported on synthetic and real-world data.


Poster Number: CSE-34

Title: An Efficient Integrated Approach to Precision Irrigation System Design for Optimal Usage Using EMO and Subsurface Water Retention Technology (SWRT)

Authors: Proteek Roy; Kalyanmoy Deb

Abstract: For growing population in today's world, water is vital for food and biomass production and it is necessary to make effective use of water. Subsurface Water Retention Technology (SWRT) has been invented for minimizing irrigation water supply and maximizing retention of water in the root zone. Irrigation scheduling and shape and placement of SWRT membranes depend on one another to achieve optimal yield. Here we have integrated water flow and nutrient transport modeling software- HYDRUS2D with an Evolutionary Multi-objective Optimization (EMO) algorithm, namely NSGA-II, to find optimal membrane geometry and placement in soil along with maximum retention of surface water. Two objectives that we optimize are 1) maximization of root water uptake efficiency (RUE) indicating the amount of water available within the soil root zone and 2) maximization of water use efficiency (WUE) indicating the available water inside the membranes. We have split irrigation system in weekly scheduling manner and rainfall information is incorporated throughout crop growing season. We have implemented an efficient approach of EMO algorithm by parallel implementation of function evaluation and recently developed efficient non-dominated sorting algorithm. Time complexity of this efficient algorithm is lower than the state-of-the-art algorithm in the worst case when statistical independence is assumed among objectives. Although this fast algorithm has little effect on the running time comapared to time taken by HYDRUS software, our theoretical investigation provides future consequences on the algorithm. Our overall results suggest that we can choose optimal scheduling strategy with corresponding SWRT technology to minimize water supply and maximize surface water retention.

This work was supported in part by BEACON Center for the Study of Evolution in Action


Poster Number: CSE-35

Title: Task Learning through Visual Demonstration and Situated Dialogue

Authors: Changson Liu; Sari Saba Sadiya; Shaohua Yang; Joyce Y. Chai

Abstract: To enable effective collaborations between humans and cognitive robots, it is important for robots to continuously acquire task knowledge from human partners. to address this issue, we are currently developing a framework that supports task learning through visual demonstration and natural language dialogue. One core component of this framework is the integration of language and vision that is driven by dialogue for task knowledge learning. This paper describes our on-going effort, particularly, grounded task learning through joint processing of video and dialogue using And-Or-Graphs (AOG).

This work was supported in part by DARPA SIMPLEX program N66001-15-C-4035


Poster Number: CSE-36

Title: Incremental Acquisition of Verb Hypothesis Space towards Physical World Interaction

Authors: Lanbo She; Joyce Y. Chai

Abstract: As a new generation of cognitive robots start to enter our lives, it is important to enable robots to follow human commands and to learn new actions from human language instructions. While grounding language to perception has received much attention in recent years, few work has addressed grounding language to action. to address this limitation, this paper presents an approach that explicitly represents verb semantics through hypothesis spaces of fluents and automatically acquires these hypothesis spaces by interacting with humans. The learned hypothesis spaces can be used to automatically plan for lower-level primitive actions towards physical world interaction. Our empirical results have shown that the representation of hypothesis space of fluents, combined with the learned hypothesis selection algorithm, outperforms a previous leading approach. In addition, our approach supports incremental learning which can serve as a basis for future life-long learning from humans.

This work was supported in part by This work was supported by IIS-1208390 from the National Science Foundation and N00014-11-1-0410 from the Office of Naval Research.


Poster Number: CSE-37

Title: Machine Learned Learning Machine

Authors: Leigh Sheneman; Arend Hintze

Abstract: In nature, organisms demonstrate the ability to both evolve and learn. At the evolutionary level, selective pressures can influence how the organisms brain is organized and how it’s components work, which shapes how the organism learns during it’s lifetime. While there are many different types of learning, here we define learning as the ability to integrate the consequences of past behaviors into future decision making. When placed in an ambiguous environment, organisms must learn what task will lead to a performance improvement. Evolutionary experiments in natural organisms would require long periods of time, making it virtually impossible to study the evolution of learning. On top of that, the neural mechanisms involved in learning must first be identified, and even then accessing the inner-workings of the brain is difficult to say the least. Recent research has been moving towards developing computational models that shed light on the natural systems in question. However, to date there has not been a system that demonstrates the ability to evolve an organisms capability to learn in it’s lifetime. To remedy that shortcoming, we invented a new type of gate for Markov Brain Networks (a form of evolvable neural network) that allows agents to incorporate feedback into action choices. We used an environment to test the agents' ability to learn over evolutionary time. Furthermore, we show that the agents also are able to incorporate their experiences into their decision making process. Thus, we have created a model that is able to change its underlying structure as well as react to its environment – a system that learns.


Poster Number: CSE-38

Title: Predicting Missing Demographic Information in Biometric Records Using Label Propagation Schemes

Authors: Thomas Swearingen; Arun Ross

Abstract: Biometric systems use biological attributes such as face, fingerprint, or iris to automatically recognize an individual. In many law enforcement applications, the biometric record of a person in the database is often supplemented with demographic data such as age, race, gender, etc. In such applications, some of the records may have missing or incorrect demographic data.  In this work, we develop a Label Propagation method to impute demographic data to partially incomplete biometric records.  The proposed method utilizes a graph-like structure to capture similarities between biometric records based on the face image, name, gender and race of individuals.  This structure is then used by the Label Propagation method to predict missing data. Experiments confirm the efficacy of the scheme in predicting missing values in biometric records.


Poster Number: CSE-40

Title: Learning with Missing Modalities via Cascaded Residual Autoencoder

Authors: Luan Tran; Xiaoming Liu; Jiayu Zhou; Rong Jin

Abstract: Aordable sensors lead to an increasing interest in acquiring and modeling data with multiple modalities. In the domain of object recognition, learning from multiple modalities has shown to signicantly improve the recognition performance. However, in practice it is very common that the sensing equipment experiences unforeseeable sensor malfunction or conguration issues, leading to corrupted data points with one or more missing modalities for learning. Most existing multi-modal learning algorithms could not handle missing modalities, and would discard either all modalities with missing values or all corrupted data. To leverage the valuable information in these corrupted data, we propose to impute the missing data given the observed information, by leveraging the relatedness among dierent modalities. While imputation has been well studied for missing at random (MAR), imputing the block-wise missing data of modalities is rarely studied. The problem is challenging because methods developed for MAR are not capable of recovering enough details for corrupted modalities, leading to a suboptimal recognition performance. In this paper, we propose a novel Cascaded Residual Autoencode (CRA) to impute missing modalities. By stacking residual autoencoders, CRA grows iteratively to model the residual between the current prediction and original data. Extensive experiments demonstrate the superior performance of the CRA on both the data imputation task and the object recognition task on the imputed data.

This work was supported in part by NGA


Poster Number: CSE-41

Title: Privacy Preserving Data Publishing for Medical Data

Authors: Ding Wang; Pang-Ning Tan

Abstract: Privacy preserving data publishing (PPDP) has attracted considerable attention in recent years due to the pressing need for publishing data without comprising users' confidential information. In this project, we examined the limitations of existing data perturbation methods for anonymizing patients' medical data. A major challenge in PPDP is to strike a balance between data privacy and data utility, the latter of which refers to the value of data after anonymization. We measure the utility of the anonymized data in terms of how well it can be used to train accurate models for classifying medical data. We argue that existing PPDP methods may significantly degrade the performance of classifiers since the perturbations were made in an unsupervised fashion, i.e., with no regards to the true class distribution of the data. To overcome this problem, we present a novel approach called surrogate learning that employs an out-of-domain feature transformation approach to transform the data into a new representation, while preserving its data utility and providing theoretically proven privacy guarantees.


Poster Number: CSE-42

Title: Discriminative Fusion of Multiple Brain Networks for Early Mild Cognitive Impairment Detection

Authors: Qi Wang; Jiayu Zhou

Abstract: In neuroimaging research, brain networks derived from different tractography methods may lead to different results and perform differently when used in classification tasks. As there is no ground truth to determine which brain network models are most accurate or most sensitive to group differences, we developed a new sparse learning method that combines information from multiple network models.


Poster Number: CSE-43

Title: A Performance Study of the Impact of Recombination and Other Evolutionary Processes on State-of-the-Art Phylogenetic Inference Methods

Authors: Zhiwei Wang; Kevin Liu

Abstract: The phylogeny, or evolutionary history, of a set of genomes is shaped by recombination acting alongside other evolutionary processes such as point mutations and genetic drift. Phylogenies are typically inferred from biomolecular sequence data using computational approaches. The most widely-used state-of-the-art methods assume that genomic loci are independently and identically distributed -- an assumption made for pure mathematical convenience -- which effectively assumes: (1) infinite recombination between loci and (2) zero recombination within each individual locus. Past studies have shown that the first assumption (i.e., no intra-locus recombination) has a relatively small impact on the accuracy of phylogeny inference when compared to other factors (e.g., evolutionary divergence). However, *both* the first and second assumptions are commonly violated in many empirical phylogenetic studies. A major open question remains: what is the impact of violations of the second assumption (infinite inter-locus recombination) upon state-of-the-art phylogenetic inference methods? To investigate this question, we conducted an extensive performance study using simulated and empirical datasets. Preliminary results confirm that the accuracy of state-of-the-art methods is degraded by violations of the assumption of infinite inter-locus recombination. We note that the state of the art tends to treat recombination as a nuisance; in contrast, we demonstrate that the genomic patterns created by recombination represent a useful signal for inference purposes, much like the patterns left by point mutations.

This work was supported in part by This work was partially supported by Grant CCF-1565719 from the National Science Foundation and by startup funds from Michigan State University (to K.L.).


Poster Number: CSE-44

Title: Multi-Task Learning with Tensor Decomposition and its Application on Geospatio-Temporal Data

Authors: Jianpeng Xu; Jiayu Zhou; Pang-Ning Tan; Lifeng Luo

Abstract: Geospatio-temporal data mining is essential to important applications in domains such as ecology, medicine and agriculture sciences. Predictions for a response variable are usually required for multiple locations, which raises the interest of learning the models for multiple locations simultaneously. Meanwhile, some of the climate phenomena are also interesting to be explored, such as climate teleconnection and El Ni˜ no. Taking these climate phenomena into the predictive framework could potentially help to improve the performance. In this paper, we propose a multi-task learning framework that builds the predictive models for multiple locations simultaneously. We incorporate tensor decomposition techniques in the framework to explore the relationship between tasks implicitly. The factors learned from tensor decomposition can be regarded as task group, timeseries group and feature group, which might represent interesting climate concepts, such as regions with teleconnections, possible climate indices, and hidden feature space. Experiments are performed on a real world climate data composed by the monthly data from over 1000 weather stations globally in more than 30 years.

This work was supported in part by This research is partially supported by NOAA Climate Program office through grant #NA12OAR4310081 and NASA Terrestrial Hydrology Program through grant #NNX13AI44G.


Poster Number: CSE-45

Title: Grounded Semantic Role Labeling

Authors: Shaohua Yang; Qiaozi Gao; Changsong Liu; Caiming Xiong; Joyce Chai

Abstract: Semantic Role Labeling (SRL) captures semantic roles (or participants) such as agent, patient, and theme associated with verbs from the text. While it provides important intermediate semantic representation for many traditional NLP tasks (such as information extraction and question answering), it is not aimed for capturing grounded semantics so that an artificial agent can reason, learn, and perform the actions with respect to the physical environment. To address this problem, this work extends traditional SRL to grounded SRL where arguments of verbs are grounded to participants of actions in the physical world. By integrating language and vision processing through joint inference, our approach not only grounds explicit roles, but also grounds implicit roles that are not mentioned in language descriptions.

This work was supported in part by DARPA SIMPLEX program N66001-15-C-4035


Poster Number: CSE-46

Title: Multi-Task Convolutional Neural Network for Face Recognition in Constrained Environment

Authors: Xi Yin; Xiaoming Liu

Abstract: Face recognition is a challenging problem due to its variations in pose, expression, illumination, etc.

It has long been considered as a single-task learning problem of extracting robust facial features. In this paper, we study face recognition on Multi-PIE using a multi-task convolutional neural network (MT-CNN). First, we consider face recognition as a multi-task problem where extracting identity feature is the main task and estimating pose, expression, and illumination are side tasks. We show that performing side tasks helps face recognition. Second, we propose a task-directed MT-CNN framework by grouping different poses to learn pose-specific identity features, simultaneously across all poses. During the testing stage, the estimated pose is used as a signal to automatically direct feature extraction. Extensive experiments on the entire dataset of Multi-PIE demonstrate the effectiveness of the proposed approach. To the best of our knowledge, this is the first paper that uses all data in Multi-PIE for face recognition.


Poster Number: CSE-47

Title: Analysis of Bounds on Hybrid Vector Clocks

Authors: Sorrachai Yingchareonthawornchai; Sandeep Kulkarni; Murat Demirbas; Eric Torng

Abstract: Hybrid vector clocks (HVC) implement vector clocks (VC) in a space-efficient manner by exploiting the availability of loosely-synchronized physical clocks at each node. In this paper, we develop a model for determining the bounds on the size of HVC. Our model uses four parameters,: uncertainty window,δ: minimum message delay,α: communication frequency andn: number of nodes in the system. We derive the size of HVC in terms of a differential equation, and show that the size predicted by our model is almost identical to the results obtained by simulation. We also identify closed form solutions that provide tight lower and upper bounds for useful special cases.Our model and simulations show the HVC size is a sigmoid function with respect to increasing; it has a slow start but it grows exponentially after a phase transition. We present equations to identify the phase transition point and show that for many practical applications and deployment environments, the size of HVC remains only as a couple entries and substantially less than n.We also find that, in a model with random unicast message transmissions, increasing n actually helps for reducing HVC size.

This work was supported in part by This work is supported by NSF CNS 1329807, NSF CNS 1318678,NSF XPS 1533870, and XPS 1533802.


Poster Number: CSE-48

Title: Hashing for Incomplete Data Source with Application of Predicting Lake Water Chemistry Data

Authors: Shuai Yuan; Pang-Ning Tan

Abstract: It is known that the real world datasets often contain a large number of features. The high dimensionality of the data makes pair-wise computation extremely expensive. to conquerer this problem, hashing based method has been introduced. Hashing is an approach to transfer data to a low dimensional representation that can preserve a certain property in the original high dimensional space. It is very popular in similarity search due to its memory and computational efficiency. Besides the high dimensionality, real world data is always noisy and contains different amount of missing values. In this poster we propose a framework that can hash the data with lots of missing values from a high dimensional space into a low dimensional representation in the meantime preserve certain characteristics.

This work was supported in part by NSF


Poster Number: CSE-49

Title: Cross-spectral Periocular Biometrics

Authors: Steven Hoffman; Muhammad Jamal Afridi; Arun Ross

Abstract: This work deals with the problem of matching ocular images of an individual across multiple spectral bands. The ocular region of the face consists of the eye - including the iris - and the surrounding skin region. The images considered in this work pertain specifically to two spectral bands: near infrared (NIR) and visible (VIS). Most iris recognition systems capture the ocular image of an individual in the NIR spectrum. However, in many legacy face databases, the ocular region is typically imaged in the VIS spectrum. In order to facilitate matching across these two modalities, we design a registration and feature extraction scheme for ocular and periocular recognition. This problem, often referred to as “heterogeneous biometrics”, also has applications in tactical scenarios where the biometric data of an individual (e.g., face or iris) is impacted by the type of sensor used for procuring data (e.g., infrared, shortwave infrared, thermal). We present experimental results to evaluate the efficacy of the proposed approach.