Human Supervision for Machine Automated Decision Making
Presented by: Baocheng Geng, Advisor: Pramod Varshney
Poster number: 32
Utilizing human expertise in addition to machine observations can be advantageous in making high-quality signal detection decisions, especially when the physical sensors are not able to collect informative data or in a critical environment that demands a high level of accuracy. In conventional detection theory, it is typically assumed that observations are collected by machines (or physical sensors) and a decision is made automatically according to the predesigned decision rule. In critical environments, the machine judgment may not be reliable and incorporating human expertise along with machine observations can be beneficial in making high-quality decisions. This study of supervision in machine automation provides insights into the design of user-focused intelligent systems and is the first step towards the development of large scale human-machine collaboration network.
Obstacle Detection and Classification with Portable Uncalibrated Patterned Projected Light
Presented by: Burak Kakillioglu, Advisor: SenemVelipasalar
Poster number: 33
Autonomous obstacle avoidance systems are instrumental for applications including assisting visually impaired people, guiding autonomous robots, and assisted driving. In this paper, we propose a novel method for autonomous obstacle detection and recognition, which employs a different type of sensor incorporating a patterned light with a camera, and provides a lightweight, cost-conscious, energy-efficient, reliable and portable solution. Experimental results indicate that grids, projected by the patterned light source, are apparent and differentiable as the sensing system is hand-carried in both indoor and outdoor environments over and towards different types of obstacles. We propose methods for exploiting these patterns for obstacle detection and classification without requiring calibration. We demonstrate that obstacles can be classified by using our novel patterned light and camera setup and employing a Convolutional Neural Network (CNN)-based approach. The proposed method provides very promising results with overall detection and classification accuracies of 93.11% and 86.06% respectively.
Actor-Critic Deep Reinforcement Learning for Dynamic Multichannel Access
Presented by: Chen Zhong, Advisor: M. CenkGursoy
Poster number: 34
We consider the dynamic multichannel access problem, which can be formulated as a partially observable Markov decision process (POMDP). We first propose a model-free actor-critic deep reinforcement learning based framework to explore the accessing policy. To evaluate the performance of the proposed accessing policy and the framework’s tolerance against uncertainty, we test the framework in scenarios with different channel switching patterns and consider different switching probabilities. Then, we consider a time-varying environment to identify the adaptive ability of the proposed framework. Additionally, we provide comparisons with the Deep-Q network (DQN) based framework, in terms of both average reward and the time efficiency.
Real-time Onboard UAV Localization without Relying on GPS Data
Presented by: Burak Kakillioglu, Advisor: SenemVelipasalar
Poster number: 35
Autonomous vehicles often benefit from the Global Positioning System (GPS) for navigational guidance as people do with their mobile phones or automobile radios. However, since GPS is not always available or reliable everywhere, autonomous vehicles need more reliable systems to understand where they are and where they should head to. Moreover, even though GPS is reliable, autonomous vehicles usually need extra sensors for more precise position estimation. In this work, we propose a localization method for autonomous Unmanned Aerial Vehicles (UAVs) for infrastructure health monitoring without relying on GPS data. The proposed method only depends on depth image frames from a 3D camera (Structure Sensor) and the 3D map of the structure. Captured 3D scenes are projected onto 2D binary images as templates, and matched with the 2D projection of relevant facade of the structure. Back-projections of matching regions are then used to calculate 3D translation (shift) as estimated position relative to the structure. Our method estimates position for each frame independently from others at a rate of 200Hz. Thus, the error does not accumulate with the traveled distance. The proposed approach provides promising results with mean Euclidean distance error of 13.4 cm and a standard deviation of 8.4cm.
Energy Efficient Occupancy Detection
Presented by: Fatih Altay, Advisor: SenemVelipasalar
Poster number: 36
This research aims to provide higher person detection rates for energy efficient vision systems.
Object Classification from 3D Volumetric Data with 3D Capsule Networks
Presented by: Burak Kakillioglu, Advisor: SenemVelipasalar
Poster number: 37
The proliferation of 3D sensors induced 3D computer vision research for many application areas including virtual reality, autonomous navigation, and surveillance. Recently, different methods have been proposed for 3D object classification. Many of the existing 2D and 3D classification methods rely on convolutional neural networks (CNNs), which are very successful in extracting features from the data. However, CNNs cannot sufficiently address the spatial relationship between features due to the max-pooling layers, and they require a vast amount of training data. In this work, we propose a model architecture for 3D object classification, which is an extension of Capsule Networks (CapsNets) to 3D data. Our proposed architecture, referred to as 3D CapsNet, takes advantage of the fact that a CapsNet preserves the orientation and spatial relationship of the extracted features, and thus requires fewer data to train the network. We compare our approach with ShapeNet on the ModelNet database and show that our method provides performance improvement especially when training data size gets smaller.
FPGA-based Neuromorphic System for Sensor Data Processing
Presented by: Haowen Fang, Advisor: Qinru Qiu
Poster number: 38
Human brain’s computational power, energy efficiency, and its ability to process real-time sensor data are attractive features for the evolving trend of the Internet of Things. Thus, drawing inspiration from the architecture of the human brain, brain-inspired computing is based on spiking neural networks (SNN). SNN consists of networks of homogeneous computing elements, i.e., neurons communicating with spikes. The sparsely distributed asynchronous events, i.e., spikes allow for a highly parallel and distributed computing architecture. Thus, neuromorphic hardware is highly desirable for embedded self-contained applications requiring real-time processing sensor data. However, there are a few challenges that limit the application of neuromorphic hardware: limited weight precision, I/O precision, training difficulty caused by the discrete nature of SNN. We developed a general workflow to mitigate these issues. To demonstrate its effectiveness, we built an FPGA-based neuromorphic system and trained a neural network to recognize sign language, and then used the proposed workflow to map it to the neuromorphic system.
Recovering secrets without reverse engineering via side-channels
Presented by: Ju Chen, Advisor: Yuzhe Tang
Poster number: 39
The information leakage through side-channels is an old topic but has been extensively re-studied recently, due to the emerging of the commercially available trusted execution environments (e.g., Intel’s SGX and ARM’s TrustZone). Since 2015, researchers have found various types of side-channels (via page tables, caches, and branch predictions) in this new setting. In 2018, researchers found another type of side-channel through the timing variance for program execution at a CPU instruction level .
The discovery of this new type of side-channel poses another research question, can we automatically correlate the leaky information with the secret program inputs? In the previous works, the authors used reverse engineering and elementary statical analysis to manually recover the protected secrets the noisy leaking messages. In this project, I proposed to use a machine learning technique to automatically recover secrets and tried to understand what the machine learning solution means for the security under the side-channel attacks
Broadband In-Phase 4-Way Equal Split Planar Power Divider/Combiner Capable of High Power Applications
Presented by: Jeremy Furgal, Advisor: Jay Lee
Poster number: 40
Phased array antenna systems are widely used in radar applications and are being researched for use in with 5G. One of the limitations of these antenna systems is the bandwidth at which they can operate. Wideband components, such as power dividers, are key in the implementation of feed networks needed for wideband phased arrays. This power divider addresses the wideband concern and also is capable of handling high power. This makes it suitable for the transmit side of the antenna network where the power is typically much higher.
Low-Profile Plasma-based Tunable Absorber for Stealth Application
Presented by: Komlan Payne, Advisor: Jay K.Lee
Poster number: 41
With the proliferation of modern frequency hopping radar tracking systems, adaptive/tunable absorbing systems are crucial for stealth technology. Microwave absorbers design is used to mitigate the radar cross section (RCS) of a target in a designated spectrum range from hostile radar signals for military and aerospace applications. The ultimate goal of this project is to investigate the feasibility of devising large scale tunable/adaptive absorbers that can maintain stable operation in harsh and dynamic electromagnetic battlefield surveillance. The use of existing mainstream technologies including PIN diodes, varactors, MEMs devices, liquid crystal polymers, and graphene require tradeoffs in terms of tuning range, cost, reliability, and linearity. In this project, we propose a low-profile tunable absorber enabled by discrete plasma-shells. Some of the advantages using discrete plasma-shells as tuning component include polarization insensitive, simple biasing scheme, wide tuning range and easy integration in many systems.
Attribute-based Object Localization
Presented by: Krittaphat Pugdeethosapol, Advisor: Qinru Qiu
Poster number: 42
Despite the recent advances in object detection, it is still a challenging task to localize a free-form textual phrase in an image. Unlike locating objects over a deterministic number of classes, localizing textual phrases involves a massively larger search space. Thus, along with learning from the visual cues, it is necessary to develop an understanding of these textual phrases and its relation to the visual cues to reliably reason about locations of described by the phrases. Spatial attention networks are known to learn this relationship and enable the language-encoding recurrent networks to focus its gaze on salient objects in the image. Thus, we propose to utilize spatial attention networks to refine region proposals for the phrases from a Region Proposal Network (RPN) and localize them through reconstruction. Utilizing in-network RPN and attention allows for an independent/self-sufficient model and interpretable results respectively.
3D Space-to-Frequency Mapping Antenna using Via-Less Composite Right/Left Handed Striplines
Presented by: Michael Enders, Advisor: Jay Lee
Poster number: 43
This work uses Composite Right/Left Handed transmission lines to reduce feed network dimensions necessary for frequency scanning antennas. Frequency scanning is implemented in a 2D planar antenna array to provide full 3D scanning of upper hemisphere. Stripline technology with inherent shielding is used to prevent electromagnetic leakage of feed network affecting radiation pattern. Through a via-less fabrication process, all this scanning capability is achieved in an integrated prototype that has been built entirely on campus. The simplicity of beam steering along with integrated passive design makes this sort of system of interest in cost-driven applications such as in automotive radars.
Learning-Based Antenna Splitting and User Scheduling in Full Duplex Massive MIMO Systems
Presented by: Mangqing Guo, Advisor: M. Cenk Gursoy
Poster number: 44
We focus on the joint antenna splitting and user scheduling problem in full duplex massive MIMO systems. By converting the original problem into a Kullback-Leibler (KL) divergence minimization problem, and solving it through a related dynamical system via a stochastic gradient descent method, we can approach the optimal solution of the original problem while the complexity is much lower, compared with the naive exhaustive search method. Our proposed algorithm can also be used to solve some other combinatorial optimization problems.
Critical Component Identification under Load Uncertainty for Cascading Failure Analysis
Presented by: Mirjavad Hashemi, Advisor: SaraEftekharnejad
Poster number: 45
As electric power grids are increasingly growing in scale and complexity, modeling and analysis of cascading failures become more critical. In this work, a new method based on distributed slack power flow is proposed for modeling and analysis of failures under load uncertainty. The developed model differs from state-of-the-art cascading failure models from two aspects: it does not require an optimal power flow solution between subsequent failures, and it alleviates the unrealistic assignment of criticality to the transmission lines that are in the vicinity of the slack bus. The proposed scheme for cascading failure analysis is implemented on a forecasted operating point of the power grid and a set of uncertainty scenarios that can occur at that operating point. The results from distributed slack cascading failure model with load uncertainty are analyzed using indices that rank components based on their appearance in cascading failure patterns. The proposed model and analysis are tested on Illinois 200-bus system, and results are discussed.
Online Design of Optimal Precoders for High Dimensional Signal Detection
Presented by: Prashant Khanduri, Advisor: Pramod Varshney
Poster number: 46
In this work, we propose a novel methodology to design optimal compression strategies (precoders) for distributed detection of high dimensional signals. We consider a wireless sensor network (WSN) which consists of multiple sensors that are spatially distributed in a region of interest (RoI) and a fusion center (FC). The sensors observe an unknown high dimensional signal and forward their observations to the FC after precoding. The sensors collect data over both temporal and spatial domains. The FC performs a binary hypothesis test based on the data received from the sensors over noisy channels. We present a technique to design optimal online linear precoding strategies in this setup in the presence of transmitting power constraints. We show analytically that the error exponents achieved by the proposed precoders are independent of the signal dimension in contrast to the state-of-the-art precoding strategies, where the error exponents decay with the signal dimension. Thus, the proposed precoder strategy performs very well even when the signal to be detected a high-dimensional signal.
On Self-localization and Tracking with an Unknown Number of Targets
Presented by: Pranay Sharma, Advisor: Pramod Varshney
Poster number: 47
We propose a scalable algorithm for cooperative self-localization and multi-target tracking. Mobile agents localize themselves and track an unknown number of targets in the presence of data-association uncertainty. We accomplish this task by modeling the probabilistic relationship between agent and target measurements alongside their state estimates using a suitably designed factor graph, which is solved using loopy belief propagation. By coupling the self-localization and multitarget tracking approaches together, our approach is capable of leveraging the statistical correlation between the agent and target state uncertainty in order to provide improved localization accuracy over time for both agents and targets. Simulation results show improvement in target and agent state estimation error, compared to the conventional approach of first localizing agents, and then using this information for target tracking.
Risk Impact Assessment of Coordinated Cyber Attacks on PMU-based State Estimator
Presented by: Sagnik Basumallik, Advisor: SaraEftekharnejad
Poster number: 48
Lack of situational awareness and improper modeling of interconnected power systems were among the most important causes of major blackouts in the past decades. Deployment of Energy Management Systems (EMS) improved situational awareness, yet additional cyber security problems were introduced. Substantial research has since been dedicated to the feasibility and consequences of False Data Injection Attacks (FDIA) and Coordinated Cyber Attacks (CCA) against power system State Estimators (SE). However, the full extent of their impacts contributing to cascading failures are not widely explored.
Moreover, attacks against Remedial Action Schemes (RAS) have not been extensively considered. In this paper, we carry out a risk impact assessment of FDIA-CCA against a PMU-based SE. A Semi-Markov Process (SMP) is used to model the probability of cyber attacks targeting Phasor Data Concentrators (PDC). To estimate the physical impacts of FDIA-CCA when RAS trigger signals are also targeted, a Distributed Slack (DS) cascading failure algorithm is developed. Extensive simulations are carried on synthetic 200 bus and 500 bus systems.
Noisy 1-Bit Compressed Sensing With Side-Information
Presented by: Swatantra Kafle, Advisor: Pramod K. Varshney
Poster number: 49
My research deals with the sparse signal reconstruction from 1-bit compressed measurements in a noisy setting. I further, improved the reconstruction performance when the receiver has access to a signal similar to the sparse signal called as side-information. This side-information is assumed to be dependent with the signal that we want to reconstruct form noisy 1-bit measurements. I looked into two different kinds of side-information: homogeneous side-information, and heterogeneous side-information. Homogeneous side-information is useful in the sequential reconstruction problem like video frame reconstruction, dynamic MRI reconstruction. Heterogeneous side-information is useful when the receiver has side-information of a different modality. We use copula functions to model the heterogeneous dependence between the side-information and the signal. To the best of the author’s knowledge, this works first to use copula-based functions capture dependence between the side-information and signal for 1-bit compressed sensing problem.
Fusion of Deep Neural Networks for Activity Recognition: A Regular Vine Based Approach
Presented by: Shan Zhang, Advisor: Pramod Varshney
Poster number: 50
The recent development of small-size, low-power, multifunctional sensors and revolutionary advances in communication and computing technologies have resulted in many applications of sensor/information fusion. One such application is human activity recognition (HAR) that can be used for health care, personal fitness, and border surveillance, etc.. We proposed a regular vine copula-based fusion of multiple deep neural network classifiers for the problem of multi-sensor based human activity recognition. We took the cross-modal dependence into account by employing regular vine copulas that are extremely flexible and powerful graphical models to characterize complex dependence among multiple modalities. Multiple deep neural networks were used to extract high-level features from multi-sensing modalities, with each deep neural network processing the data collected from a single sensor. The extracted high-level features were then combined using a regular vine copula model.
Wind Energy Intermittency Mitigation by Customer Response to Real-Time Pricing of Electricity
Presented by: Wolf Peter Jean Philippe, Advisor: Sara Eftekharnejad
Poster number: 51
In the effort to produce cleaner energy, more renewable resources are added worldwide to existing power systems in recent years. Among those, wind energy resources are becoming more viable as alternatives to fossil fuels power plants. Wind energy continues to become more affordable and cost-efficient due to technological advancements in recent decades. However, wind energy suffers a critical drawback: intermittency. The intermittency makes meeting real-time consumer demands a challenging task. Hence, utilities are forced to have standby power resources known as spinning reserves. While expensive, those reserves are often not utilized.
This research focuses on reducing the cost of spinning reserves in the power system while considering the intermittency of wind energy resources. To achieve this objective, a demand response (DR) program is proposed. The DR model is based on real-time pricing (RTP) of electricity, where customers receive incentive payments when reducing their energy consumption in a period of high electricity price, or shifting some of their consumption from high price to low price periods.
The results obtained prove that the proposed DR helps reduce the cost of spinning reserves in power systems where high wind energy penetration is considered. Therefore, the wind energy penetration can be increased, while power system utilities can securely meet real-time demand.
Coverage analysis for UAV-assisted mmWave IoT Cellular Networks
Presented by: XUEYUAN WANG, Advisor: Mustafa Gursoy
Poster number: 52
In this work, we jointly consider the downlink simultaneous information (SWIPT) and energy transfer and uplink information transmission in unmanned aerial vehicle (UAV)-assisted millimeter Wave (mmWave) Internet of Things (IoT) cellular networks, in which the UE locations are modeled as a Thomas cluster process or Mat\’ern cluster process. Through numerical results, we investigate the impact of key system parameters on the performance, which gives us insight on how to design the system parameters to obtain optimal performance of this practical system model.
Autonomous Waypoints Planning and Trajectory Generation with Deep Reinforcement Learning
Presented by: Yilan Li, Advisor: QinruQiu
Poster number: 53
Safe and effective operations for multi-rotor unmanned aerial vehicles (UAVs) demand obstacle avoidance strategies and advanced trajectory planning and control schemes for stability and energy efficiency. However, existing work does not consider the relationship between optimal control and the surrounding environment. To solve those problems in one framework analytically is extremely challenging when the UAV needs to fly large distance in a complex environment. To address this challenge, we purpose a two-level strategy that ensures a globally optimal solution can be obtained. At the higher-level, deep reinforcement learning is adopted to select a sequence of waypoints which lead the UAV from its current position to the destination. At the lower-level, an optimal trajectory is generated between each pair of adjacent waypoints. While the goal of trajectory generation is to maintain the stability of the UAV, the goal of the waypoints planning is to select waypoints with the lowest control thrust consumption throughout the entire trip while avoiding collisions with obstacles.
Energy Efficiency Optimization in UAV-Assisted Communications and Edge Computing
Presented by: Yang Yang, Advisor: M. Cenk Gursoy
Poster number: 54
We introduce uplink NOMA technique for offloading computation tasks in a UAV-based communication network.
By guaranteeing the computation latency constraints, we formulate a total energy consumption minimization problem while jointly optimizing the power allocation and the trajectory of UAV.
Based on a three-step optimization approach, we first obtain the results of the optimization problem under a giving trajectory of UAV; then we optimize the trajectory according to the previous conclusions.
Efficient Human Activity Classification from Egocentric Videos Incorporating Actor-Critic Reinforcement Learning
Presented by: Yantao Lu, Advisor: Senem Velipasalar
Poster number: 55
In this work, we introduce a novel framework to significantly reduce the computational cost of human temporal activity recognition from egocentric videos while maintaining the accuracy at the same level. We propose to apply the actor-critic model of reinforcement learning to optical flow data to locate a bounding box around the region of interest, which is then used for clipping a sub-image from a video frame. We also propose to use one shallow and one deeper 3D convolutional neural network to process the original image and the clipped image region, respectively. We compared our proposed method with another approach using 3D convolutional networks on the recently released Dataset of Multimodal Semantic Egocentric Video. Experimental results show that the proposed method reduces the processing time by 36.4% while providing comparable accuracy at the same time.
Distributed Resource Allocation in Wireless Networks via Deep Reinforcement Learning
Presented by: ZIYANG LU, Advisor: M. Cenk Gursoy
Poster number: 56
In wireless communication networks, the spectrum is a scarce resource that needs to be effectively utilized as mobile data traffic over wireless networks grows at an unprecedented rate. Prior research efforts have addressed efficient spectrum allocation via optimization algorithms. However, optimization methods usually require complete knowledge of the network and are difficult to be implemented in dynamic environments.
Recently, deep reinforcement learning (DRL) has been widely applied to solving diverse, dynamic control problems. Here, we propose a DRL-based method of allocating limited resources (spectrum and power) to maximize specific wireless network utility objectives such as the sum transmission rates or the fairness among the users in the network. It is worth noting that there is no information exchange between the users once the deep neural network is well trained, which makes this algorithm scalable becomes practical in large-scale networks.
Human Activity Classification Egocentric Video and Inertial Measurement Unit Data
Presented by: Yantao Lu, Advisor: Senem Velipasalar
Poster number: 57
Many methods have been proposed for human activity classification, which rely either on the Inertial Measurement Unit (IMU) data or data from static cameras watching subjects. There has been relatively less work using egocentric videos, and even fewer approaches combining egocentric video and IMU data. Systems relying only on IMU data are limited in the complexity of the activities that they can detect. In this work, we present a robust and autonomous method, for fine-grained activity classification, that leverages data from multiple wearable sensor modalities to differentiate between activities, which are similar in nature, with a level of accuracy that would be impossible by each sensor alone. We use both egocentric videos and IMU sensors on the body. We employ Capsule Networks together with Convolutional Long Short Term Memory (LSTM) to analyze egocentric videos, and an LSTM framework to analyze IMU data, and capture temporal aspect of actions. We performed experiments on the CMU-MMAC dataset achieving overall recall and precision rates of 85.8% and 86.2%, respectively. We also present results of using each sensor modality alone, which show that the proposed approach provides 19.47% and 39.34% increase in accuracy compared to using only ego-vision data and only IMU data, respectively.
A Simulation Framework for Fast Design Space Exploration of Unmanned Air System Traffic Management
Presented by: Ziyi Zhao, Advisor: Qinru Qiu
Poster number: 58
With the rapid increase of the UAV applications in delivery, surveillance, and rescue mission, there is an urgent need for UAV traffic management that ensures the safety and timeliness of the missions. Accurate and fast UAV traffic prediction and resource usage estimation is a key technique in the traffic management system. It allows us to 1) evaluate different mission schedules or 4G/5G resource allocation schemes in a short period; and2) adjust mission control policy in real-time. A good prediction model should not only consider the UAV mission information, but also the environment information such as weather, 4G/5Gbase station distribution, etc. Its complexity is much beyond the traditional analytical approach. We propose to solve this problem by using machine learning. The model has a combined of convolutional neural network (CNN) and recurrent neural network (RNN). The inputs are multi-channel time-varying streams, such as weather map, cellular network usage map, geographical constraints, and UAV launching/landing information, etc. The outputs will be predicted UAV conflict probability map and 4G/5G channel congestion map. Compression and acceleration of the model will also be studied for real-time prediction.