Undergraduate Projects
A Distributed Wireless System of Water-Level Sensors for Real-time Measurement of Street-level Flooding
Alexandra Du, Alex Kaplan, Kaiwen Wu, Neil Seoni, Alfonso Morera, Nicholas Lester, Nicole Tan, Leonardo Duenas-Osorio, Devika Subramanian, Gary Woods
Preparing coastal cities for future hurricanes like Harvey or for chronic intense rainfall events, requires accurate modeling of flood-water flow over large areas covered by urban streets, buildings, parking lots, and other rapidly evolving built environment features. However, accurate real-time modeling at a city scale is infeasible computationally, and requires a large data set of empirical flood-water measurements over a large urban area for model calibration and validation. In order to obtain this data, Houston Solutions Lab is supporting our development of a scalable network of wireless water-level sensors that can be deployed around coastal cities like Houston. Each sensor node can be placed at or near a street, reports wirelessly to the cloud, does not require power from the grid, and will cost at most a few hundred dollars. Our proof of concept system, comprised of four nodes around the Rice campus, measured street flooding on campus roads in December 2018. The system has a web-based interface that can be read out in real time from any browser. When deployed at scale, our system can be used to support decision making for traffic routing and evacuation, both for citizens as well as emergency officials.
Accurate prediction of boundaries of high resolution topologically associated domains (TADs) in fruit flies using deep learning
John Henderson, Vi Ly, Shawn Olichwier, Pranik Chainani, Yu Liu, Ph.D., Benjamin Soibam, Ph.D.
Genomes are organized into self-interacting chromatin regions called topologically associated domains (TADs). The boundaries of TADs are conserved, regulate gene expression and have been linked to several diseases. Even though detection of TADs boundaries are important and useful, there are experimental challenges in obtaining high resolution TADs location. Here, we present computational prediction of TADs boundaries from high resolution Hi-C data in fruit flies. By extensive exploration and testing of several deep learning model architectures with hyper-parameter optimization, we show that a unique deep learning model consisting of three convolution layers followed by long short-term memory layer achieves an accuracy of 96%. This outperforms feature based model’s accuracy of 91% and an existing method’s accuracy of 73-78% based on motif TRAP scores. Our method also detects previously reported motifs such as Beaf-32 which are enriched in TADs boundaries fruit flies, but also several unreported motifs.
ASTRO: Drone-Based Person Re-Identification
Matthew Mutammara, Ronaldo Sanchez, Xintong Liu, Varun Suriyanarayana, Logan Lawrence, Eduardo Berg, Luke Zhang
Person reidentification is a widely studied computer vision problem because of its complex nature and broad scope. Most study in this field however, has revolved around the case when the camera itself is static. Although there has been some effort to perform person reidentification on mobile platforms this is still a growing area of research. This task is complicated not only by the quality of captured data but also the power and computational resources available on a mobile platform. In our project we attempt to build a drone that can re-identify a target person with some known distinctive feature and perform a pre-determined task such as landing or pursuing this target.
Axon Mobile – Wireless Neural Recorder
Aidan Curtis, Sophia D’Amico, Andres Gomez, Benjamin Klimko, Zhiyang Zhang, Dr. Nitin Tandon
Intracranial and extracranial EEG recordings have applications in clinical epilepsy treatment and neuroscience research. Existing neural recording system solutions are wired and require patients to remain in the hospital for weeks at a time. This process is expensive and highly likely to cause infection and stress. We are developing a wireless neural recording system to help eliminate these issues. The system has a front-end interface chip that performs signal conditioning on the neural data, a wireless link to a receiver, a novel data compression algorithm, and a power management system. The final product will be a form-factor system that can be implemented in test animals for experimentation purposes.
CHARIoT: Energy Harvesting Wireless Sensor Node for IoT
Robby Flechas, Jennifer Hellar, Nathaniel Morris, Rachel Nguyen, Brady Taylor, Robyn Torregrosa, Gene Frantz, Ray Simar, Erik Welsh
With the number of Internet of Things (IoT) devices in the world increasing drastically every year, the demand for generic IoT platforms grows with them. There are many applications for this technology, but there are currently few hardware devices that can self-power, transmit, process, and sense a variety of parameters with little modification. We have taken the initial steps toward creating a generalized hardware and software architecture to support the sensor needs of these devices.
Chiclet: Team DISSECT’s Bite-sized Microcontroller
Ronaldo Sanchez, Ray Simar
The goal of the Chiclet project is to produce a versatile printed circuit board with the smallest area possible while maintaining a high number of input/output ports on the board. We are planning on printing the boards before the end of the Spring 2019 semester. The microcontroller is 1.5″x0.8″ and features 36 GPIO pins.
Combating Power Amplifier Distortions using Neural Networks
Chance Tarver, Aryan Sefidi, Liwen Jiang, Cynthia Chen, Jenny Penaloza, Chuang Yu, Sam Li, and Joseph Cavallaro
Using a novel, neural-network (NN) method, we demonstrate digital predistortion (DPD) to combat the nonlinearities in power amplifiers (PAs) which limit the power efficiency of mobile devices, increase the error vector magnitude, and cause poor spectral containment. Traditionally, DPD is done with polynomial-based methods that use an indirect-learning architecture (ILA) which can be computationally intensive, especially for mobile devices, and overly sensitive to noise. Our approach using NNs avoids the problems associated with ILAs. Moreover, it can leverage NN accelerators, which are becoming more commonplace in processors, and has less running complexity than a polynomial based predistorter. The NN DPD effectively learns the unique PA distortions, which may not easily fit a polynomial-based model, and hence outperforms comparable polynomial-based DPDs.
Fast, Accurate, Low-cost QR-based Camera Calibration and Object Localization System for Indoor Tracking
Spencer Chang, Ishan Mehta, Fasai Phuathavornskul, Napas Udomsak, Martin Zhang, Tianyi Zhang
We propose a low-cost indoor localization and navigation system based on QR codes. Our system consists of automated camera calibration and object localization. QR corner points are used as feature points for estimate of camera-camera and camera-object transformations. In the calibration phase, QR codes are strategically placed around the indoor space and camera poses or extrinsic parameters are estimated by reconstructing the camera vision graph. During the object tracking phase, the cameras are able further compute the absolute pose of a moving object with visible QR markers. We will validate the feasibility of the system by performing experiments on rendered and real-life scenes.
Ghost Rider
Collin Allen, Anirudh Kuchibhatla, Nick Glaze, Michael Angino, Ray Simar
Our team is focusing on autonomous driving development, currently implementing autonomous capabilities on an RC motorcycle. At our current stage of development, we are able to control the motorcycle using an on-board arduino and circuitry. This has enabled us to have the motorcycle follow a pre-programmed path that can be uploaded into the motorcycle.
Optical Design for Motion Compensation in Wearable Devices
Chiraag Kaushik, Belviane Songong, Vivek Boominathan, Ashok Veeraraghavan, Ashutosh Sabharwal
Wrist based Photoplethysmography (PPG) sensors can potentially be used for robust continuous monitoring of heart rate, heart rate variability, SpO2, and other vital signs. However, a major barrier to continuous monitoring is reliable and accurate measurement of vital signs in the presence of motion artifacts. Current wearable devices utilize a single photodetector to measure changes in intensity over time, making them very susceptible to noise from motion. We propose an optical design for improved spatial information capture. Our design uses a micro-lens array (MLA) in parallel with a CMOS sensor to obtain a spatial mapping of the skin. This spatial mapping can then be used to track translational motion of the device across the skin and improve the resilience of the acquired PPG signal to motion.
Physical Skin Phantoms to Replicate Optical Properties of Human Tissue
Hannah Andersen, Aarohi Mehendale, Mary Jin, June Chen, Ashutosh Sabharwal, Ashok Veeraraghavan
The Wearables lab at Rice University aims to develop optical medical devices in order to create non-invasive diagnostic, monitoring and therapeutic applications of light. These prototypes need to be tested against known parameters. As these tests cannot be conducted directly on human tissue, the need for a tissue phantom arises. Our project focuses on producing these optical skin phantoms and characterizing them based on their optical properties.
Project Confidence: MousePass
Nickolas Chen, Timothy Goh, Amanda Lu, Michael Sprintson, Dru Myerscough, Ricky Simar
Every year, millions of internet users are victims of online identity theft. Current password infrastructure combined with a lack of culture around password security has exacerbated this problem worldwide. MousePass is a unique solution that utilizes mouse movement as a biometric to verify a user’s identity in order to prevent fraudulent access to their account. Studies show that an intruder moves his or her mouse in a significantly different pattern than someone responding truthfully on forms. MousePass takes advantage of this fact to provide an accessible yet secure interface for verification.
Project Pancreas: Prototyping Affordable Insulin Pumps
Hamza Rahim, Huzaifah Shamim, Justin Cheung, Oluwatonamise Akerele, Qixuan Yu, Shannon McGill, Paz Zait-Givon, Daniel DeSalvo, Ray Simar
Insulin pump therapy offers many health and quality of life benefits to insulin dependent diabetics. Unfortunately it is expensive and thus unavailable to many who would benefit from it. We are exploring the possibility of building an affordable modern insulin pump from off-the-shelf components along with leveraging the open source diabetic community.
Unconventional nanophotonics with non-Hermitian physics
Alex Hwang, Chloe Doiron, Gururaj Naik
From the early formulations of quantum mechanics, physical systems were always assumed to be “Hermitian,” meaning that they do not exchange energy with their environment. Considering only Hermitian systems places limits on the physical phenomena that can occur. Designing non-Hermitian systems with controlled energy exchange with the environment can produce rich new physical phenomena. For example, non-Hermitian photonic systems can exhibit ultrafine sensing and unidirectional light propagation. However, non-Hermitian photonics has not been demonstrated in the nanometer-scale, a regime with important technological implications in information technology, medicine, and energy. In this talk, we highlight our progress on designing and implementing non-Hermitian nanophotonic systems and their potential applications in areas such as sensing.
Vignette
Robin Kim, Daniel Vadasz, Carl Henderson, Ayush Chapagain, Gary Woods, Lin Zhong, and Venu Vasudevan
In this current data-driven age, companies are seeking methods to gather customer information
regarding their products in hopes of replacing the traditional online survey approach. The purpose of this project is to build a video capture system to assist companies desiring information on how often the user interacts with certain products. To incentivize having this product in the customer’s home, the company would fully cover the cost of the device as well as offer some form of financial compensation to the customer whether it be in cash or discounts on the company’s products. To differentiate this product from other smart home devices, this device will allow the user to thoroughly control all data leaving the device. Any object recognition computation should be done locally on the device to give the user ease of mind that the customer’s raw video data is not being sent to the company’s remote server without the user’s consent. Additionally, any object recognition computation done on the device will be summarized as a text file and provided to the user directly allowing the customer to control what data is allowed to leave the device.
Graduate Projects
A 562F2 Physically Unclonable Function (PUF) with a Zero-Overhead Stabilization Scheme
Dai Li, Kaiyuan Yang
Internet of Things (IoT) devices bring a growing demand for secure, low-cost, and low-cost secret key and ID storage solutions. Physically unclonable functions (PUFs) are one of the most promising alternatives to conventional non-volatile memory (NVM) solutions, which usually incurs extra cost and are vulnerable to physical attackers. In recent years, significant research efforts have been devoted to improving the stability of PUF responses/keys across PVT variations while maintaining the low cost and low power benefits. A 562F2 PUF cell, which can be reconfigured to a different PUF cell locally for lossless stabilization with zero area overhead is implemented in 65nm CMOS. It features: (1) a 0.00182% native bit error after lossless stabilization and 0.44% across the military temperature range (-55 to 125°C); (2) a low-cost stabilization scheme combining a reconfigurable PUF cell with a body-bias-based V/T emulation method, which improves bit error rate (BER) by 149 times and unstable bits by 120 times; (3) sub-threshold operation with only 0.062fJ/bit core energy.
A mm-Sized Wireless Implantable Medical Device with Magnetic Power Transfer for Neural Applications
Zhanghao Yu, Kaiyuan Yang
Nowadays, power delivery is still a fundamental challenge for miniature implantable medical applications. The interfaces of bio-electronics devices with wireless powering are potentially longer lasting and less invasive than battery powered or wired ones. Compared to energy harvesting via electromagnetic or ultrasound, the magnetic demonstrates negligible tissue absorption, significant smaller wavelength and much less impedance mismatch between air, bone and tissue [Guduru et al.], which allow more effective power delivery to implants deep inside tissue. A current state-of-the-art fabricates a in vivo magnetic wireless neural stimulator using novel material [Wickens et al.]. However, it is ineffective, uncontrollable and inefficient due to the lack of the specialized functional circuitry in receiver. As a result, this device cannot be applied for high-power stimulation. The objective of this project is designing the first integrated wireless implantable receiver using magnetic power transfer with mm-sized area. An improved matching network and power management integrated circuit will be implemented to enable effective output power while maintaining sufficient energy harvesting efficiency. In the future, this type of devices can be widely used for neural stimulation or physiological monitoring to treat some clinical diseases.
A strongly correlated material for tunable metasurfaces
Weijian Li, Cameron Gutgsell, Gururaj Naik
We demonstrate intensity dependent optical response in the visible for a strongly correlated material, 1T-TaS_2. Using this tunable material, we show the intensity-dependent diffraction of a meta-grating device useful for imaging, display and sensing technologies.
A Temperature-dependent Neural Oscillator in the Aboral Region of Hydra vulgaris
Weijian Li, Cameron Gutgsell, Gururaj Naik
We demonstrate intensity dependent optical response in the visible for a strongly correlated material, 1T-TaS_2. Using this tunable material, we show the intensity-dependent diffraction of a meta-grating device useful for imaging, display and sensing technologies.
A Trip Through the History of Wi-Fi: from 11b to 11ax
Vinicius Da Silva, Edward Knightly
The design of Medium Access Control (MAC) policies for Wi-Fi is a key step to achieve the promised multi-fold data rate gains of advanced physical layer techniques, such as Multi-User MIMO and OFDMA. At the same time, the performance of a MAC policy is highly dependent on the traffic being served by the network, with particular results for closed-loop TCP transmissions – the most common internet traffic type. Current experimental platforms for wireless research focus on the physical layer aspects and lack the capacity to run experiments with real-time network traffic, which hinder the development of novel MAC policies that explore and leverage characteristics of real user traffic. We present a novel low-cost experimental platform for evaluation of advanced Wi-Fi MAC policies under real network traffic in commodity devices. Building upon the current understanding of the physical layer design and a multitude of previous experimental work, we emulate the behavior of a wireless network by shaping the traffic in a high-speed wired LAN according to the theoretical over-the-air transmission times of Wi-Fi. This allows for the implementation of new highly flexible MAC protocols and evaluation of performance under real traffic. In this demo users can experience how each evolution step in the Wi-Fi technology has improved the network performance – with emphasis to the user experience – and take a sneaky peek into the power of the next generation of Wi-Fi.
Nishant Mehrotra, Ashutosh Sabharwal
Millimeter and terahertz imagers are generally built with the assumption that the targets to be imaged are present in line-of-sight paths to the imagers. However, multiple paths to the target may exist in actual, urban environments with strong scatterers around the region of interest; and imagers typically do not exploit the additional signal energy present in them. In this work, we aim to improve their performance by exploiting the signal energy present in the multipath.
Yuning Wang, Sakib Hassan, Frank Yang, Gururaj Naik
Light sources with specific spatial and temporal properties are required for many applications in our daily lives, including telecommunications, medicine, and scanning. However, in order to achieve these specifications by traditional light sources such as lasers with different filtering apparatuses, sources can become bulky and inefficient. Thus, we seek to design a novel light source that is both compact and efficient. A metasurface is an excellent choice for this purpose. A metasurface is an optical device composed of many optical scatterers that allows for control of the light emitted from a gain material. Additionally, we require a light emitting material with high optical gain. Perovskites have been considered as a promising light emitting material due to their excellent properties, such as large absorption coefficients, broadband bandgap tunability, high photoluminescence quantum efficiencies, long-distance charge transport, and low-cost fabrication. Here we aim to build a proof-of-concept metasurface using perovskites materials to achieve meta-light emission. We will investigate the optical properties of perovskites by different measurement methods, like reflection, transmission, absorption, photoluminescence, and optical gain using pump-probe. Also, using finite difference simulations of the Maxwell Equations, we will design and optimize perovskite metasurfaces from our measured data. Finally, we will fabricate the optimized device, demonstrating its potential as a compact, tunable, and efficient light source.
Daniel LeJeune, Richard G. Baraniuk, Reinhard Heckel
Algorithms often carry out equally many computations for “easy” and “hard” problem instances. In particular, algorithms for finding nearest neighbors
typically have the same running time regardless of the particular problem instance. In this paper, we consider the approximate k-nearest-neighbor
problem, which is the problem of finding a subset of O(k) points in a given set of points that contains the set of k nearest neighbors of a given query point.
We propose an algorithm based on adaptively estimating the distances, and show that it is essentially optimal out of algorithms that are only allowed to
adaptively estimate distances. We then demonstrate both theoretically and experimentally that the algorithm can achieve significant speedups relative to
the naive method.
Yang Zhao, Pengfei Xu, Yingyan Lin
There has been a tremendous need for Deep Neural Network (DNN) accelerators to bring powerful yet power hungry DNNs to resource-constrained or time-critical applications. However, designing DNN accelerators is a non-trivial task because (1) powerful DNNs have millions of parameters and billions of operations and (2) the design space of DNN accelerators is large due to the numerous design choices of dataflows, computation resources, memory size and hierarchy, etc. In this work, we propose an analytical model framework that can formulate DNN accelerators’ energy, throughput, and latency as a function of hardware parameters, DNN models’ parameters, unit energy cost and delay of computation and various memory accesses. Validation shows that our models’ predicted performance matches well with that of chip-measured ones. We further use this model framework to develop a cooperative algorithm and hardware co-design technique to aggressive reduce the energy cost of DNN accelerators.
Qingyue Liu, Peter Varman, Dan Luo
This project investigates the design space for distributed storage systems based on emerging storage and networking technologies. Recent byte-addressable non-volatile memory (NVRAM) like Intel’s 3D XPoint together with fast network fabrics like RDMA, provide opportunities for creating a new class of scalable high-throughput data management systems with low latency and strong consistency guarantees. In this research we explore the design of Cheetah, a distributed storage system that aims to support durability, multiple granularity access (cache lines and blocks), and consistency of storage operations. Cheetah exploits DIMM-attached NVRAM and hybrid RDMA primitives to achieve microsecond-level latency and high throughput.
Boqiang Fan, Behnaam Aazhang
Ultrasound neuromodulation has attracted much attention from researchers all over the world in recent years. Many experiments have shown that with ultrasound stimulating neurons in the brain, the neuronal firing will be modulated and the behavior of subjects will change as well. Scientists have proposed to apply such a new technique to the treatment for neuropsychiatric disorders, e.g. depression or obsessive compulsive disorder, by focusing the ultrasound power into corresponding brain regions. Stimulation signal patterns need to be optimized for treatment, which requires computational evaluation before in vivo experiments. In this research, we aim to computationally evaluate the intracranial propagation of focused ultrasound in brain tissues, to find potential problems with current focused ultrasound protocols and achieve neuromodulation with very high spatial resolution. Simulation results of ultrasound propagation in brain tissue constructed from brain scan data is presented. Future works will focus on further improving the ultrasound transducer and the stimulation signal pattern for modulating brain activities.
Decentralized Coordinate-Descent Data Detection and Precoding for Massive MU-MIMO
Kaipeng Li, Oscar Castaneda, Charles Jeon, Joseph R. Cavallaro, Christoph Studer
Massive multiuser (MU) multiple-input multiple-output (MIMO) promises significant improvements in spectral efficiency compared to small-scale MIMO. Typical massive MU-MIMO base-station (BS) designs rely on centralized linear data detectors and precoders which entail excessively high complexity, interconnect data rates, and chip input/output (I/O) bandwidth when executed on a single computing fabric. To resolve these complexity and bandwidth bottlenecks, we propose new decentralized algorithms for data detection and precoding that use coordinate descent. Our methods parallelize computations across multiple computing fabrics, while minimizing interconnect and I/O bandwidth. The proposed decentralized algorithms achieve near-optimal error-rate performance and multi-Gbps throughput at sub-1 ms latency when implemented on a multi-GPU cluster with half-precision floating-point arithmetic.
Deep feature learning in human sensing data
Boning Li, Akane Sano
We train deep networks to automatically extract physiological and biobehavioral features from raw data collected by wearable/mobile sensors to help understand and improve people’s health and wellbeing.
Deep generative model for graph structured data
Kuida Liu, Santiago Segarra, Reinhard Heckel
Graph structure data which are non-Euclidean are universal of real life, like social relation network and knowledge network. Other than the data in Euclidean able to be presented as vector or matrix, like speech or image signal, graph data are composed of nodes and edges between two nodes. The generative model for graph structure data comes into notice of research because of the need of big data and prosperity of deep learning. The first generative mode is introduced by Erdos but the strongly independent character makes the model lost its practicability. The rising and successful generative models in the framework of deep learning, like VAE and GAN, are hard to be applied into this task because they require a strict data structure. I expect to give a learnable generative model for Graph structure data, which provides methods to express probabilistic dependencies in nodes and edges. By applying geometric deep learning methods, embedded graph data also show compatibility to high dimension data. In the research, I will cover different graph generative model, like sequence generative model and auto-encoder on graph. By comparing advantages and combination, I expect a high-performance and robust algorithm for data generative model.
Defense against the dark arts of neurological diseases: using electromagnetic waves to stimulate deep brain regions non-invasively
Fatima Ahsan and Behnaam Aazhang
When you have a neurological disease like Parkinson, you feel that the pleasures of your life are gradually being taken away from you. You can’t eat, drive or travel alone because your hands and feet are constantly shaking. Sometimes while walking you might want to make a sharp turn but you will freeze, creating a hazard for losing balance and falling. Of course there are medications that can help reduce tremors but not only do they come with significant side-effects like drowsiness but since Parkinson’s gradually worsens with time, the medications stop having an effect. A natural question to ask is that is there any other treatment? Turns out there is one solution that works perfectly in eliminating tremors and returning the quality of life of a Parkinson’s patient but it comes with the side-effects of brain infection and stroke. This solution is called deep brain stimulation and it is highly invasive. Because two metal electrodes are surgically implanted deep inside the brain and electric pulses are sent to them with the help of wires to stimulate the motor control center deep inside the brain. My project is to stimulate this deep motor control section with the help of electromagnetic waves without having to stimulate it by surgically implanting metal electrodes in sensitive brain tissue.
Design & Implementation of Private Wearable Speakers
Mojtaba Safaei, Lin Zhong
Audio is an important modality of output for human-computer interaction, especially on mobile computers. Yet due to its broadcast nature, audio output is not private unless it is delivered in an obtrusive way, i.e., through on-ear or in-ear headsets. Therefore, this project asks this question: is it possible to deliver audio signal in a way that is both private and non-obtrusive? Being private, the audio signal should not be intelligible to humans that are beyond the intimate space, i.e., about 50 cm from the user. Being non-obtrusive, it should not require any parts in or on the user ears. To answer this question, this project explores several complementary technologies. First, we investigate the possibility of delivering acoustic signal from a proximity of the ear but without direct contact with it. For example, the hardware can sit on the shoulder or a Glass. This proximity allows us to reduce the output power and improve privacy. Second, we leverage acoustic beamforming to focus the output signal on the intended user ear to further reduce output power and improve privacy. Third and most importantly, we leverage binaural hearing to split an intended acoustic signal into two channels, one to each user ear. The signal will be split in a way that the user that hears both channels will “hear” the intended signal. We will investigate multiple ways to split the signal, considering how human binaural hearing works. Finally, we will further investigate the design of the audio output signal. The most challenging and useful audio output would be speech. Other than speech, we will also consider less challenging ways of audio output of limited usage, e.g., auditory icons. We will carry out the project with both simulation and prototyping. For acoustic beamforming, we will leverage commercial-off-the-shelf speaker arrays of small form factors. For signal splitting, we will leverage existing audio signal processing/editing tools such as MatLab or Python.
DiMuS: Multi-scale graph theoretic analysis of brain dynamics of a language task
Sudha Yellapantula, Nitin Tandon, Behnaam Aazhang
We humans are remarkably gifted in our use of language. Unfortunately, millions are ravaged with language disorders, many of which remain irremediable. A prerequisite to develop new treatments requires a much deeper understanding of the underlying neural patterns than is currently available. This work develops a new data driven analysis framework, named DiMuS (Directed Information Multi-Scale), to further our understanding of spatial, temporal and spectral brain dynamics during a language task. Electrocorticography (ECoG) data was acquired from intracranial electrodes implanted in epilepsy patients. In previous studies, task dependent high gamma power responses were observed in ECoG data, however, the underlying connectivity patterns are still unclear. The DiMuS framework has three main parts. Firstly, causality between brain regions was measured by Directed Information (DI) in a model free
manner. Secondly, time-varying brain graphs were generated and a multi-scale graph theoretic analysis extracted novel spatio-temporal dynamics at a coarse, intermediate and fine scale. Finally, the relationship between observable electrode power responses and the underlying information flow was examined. This is the first known work to find strong statistical correlations between the time-series of high gamma power responses, and the time-series of total information entering that brain region. Spectral properties of information flow in various frequency bands were also studied. The spectral results showed that the information flow associated with high gamma power responses was concentrated predominantly in the theta or beta bands, depending on the specific brain region. DiMuS can provide fresh perspectives in many neural tasks, paving the way to benefit people with language disorders.
Drug Repurposing Discovery with Graph Convolutional Networks
West Zhican Chen, Dr. Santiago Segarra, Dr. Xiaoqian Jiang
Graph Convolutional Network(GCN) is an implementation of Convolutional Neural Network(CNN) on irregular structured graph data aimed at Representative Learning task. Compared to traditional shallow embedding approaches such as Factorization based approaches and Random Walk approaches, GCN performs much better in many senses and is thus considered a powerful tool for Representative Learning. However, limited attention was paid to GCN’s performance towards multi-modal graph dataset, i.e. graph network with many different node types. The objective of this project is to examine and compare the efficacy of different variants of GCN (e.g., complex network structure vs. complex node features) with regard to drug repurpose discovery related applications. Our investigation will be based on the dataset from UT Health Science Center at Houston.
Dual Dynamic Inference: Enabling Multi-Grained and More Controllable Inference Adaptivity in Complex CNNs
Yue Wang, Jianghao Shen, Tan Nguyen, TingKuei Hu, Richard Baraniuk, Zhangyang Wang, Yingyan Lin
State-of-the-art convolutional neural network yield record-breaking predictive performance, yet at the cost of high inference costs, that prohibits their widely deployments in resource-constrained IoT applications. We propose a dual dynamic inference (DDI) framework that highlights the following aspects: 1) we integrate both input-dependent and resource-dependent dynamic inference mechanisms in one unified framework, to fit the varying IoT resource requirements in practice; 2) we propose a novel multi-grained learning to skip (MGL2S) approach for input-dependent inference, that allows simultaneously for layer-wise and channel-wise skipping and thus enables superior flexibility; 3) we extend DDI to complex CNN backbones such as DenseNet, and show that DDI can be applied towards optimizing any specific resource goal, such as inference latency or energy cost. Extensive experiments demonstrate the superior inference accuracy-resource achieved by DDI, as well as the flexibility to control such trade-offs, over existing peer methods.
Enabling a “Use-or-Share” Framework for PAL-GAA sharing in CBRS Networks via Reinforcement Learning
Chance Tarver, Matthew Tonnemacher, and Joseph Cavallaro
By implementing reinforcement learning aided listen-before-talk (LBT) schemes over a Citizens Broadband Radio Service (CBRS) network, we increase the spatial reuse at secondary nodes while minimizing the interference footprint on
higher-tier nodes. The FCC encourages “use-or-share” policies in the band across the PAL-GAA priority tiers by opportunistically allowing the lower-priority GAA nodes to access unused higherpriority PAL spectrum. However, there is currently no mechanism to enable this cross-tier spectrum sharing. We propose and evaluate LBT schemes that allow opportunistic access to PAL spectrum. We find that by allowing LBT in a two carrier, two eNB scenario, we see upwards of 50% user perceived throughput (UPT) gains for both eNBs. Furthermore, we examine the use of Q-learning to adapt the energy-detection threshold (EDT), combating problematic topologies such as hidden and exposed nodes. With merely a 4% reduction in primary node UPT, we see up to 350% gains in average secondary node UPT when adapting the EDT of opportunistically transmitting nodes.
End-to-end Learned Phase Masks for Extended Depth of Field Imaging with a Large Aperture
Shiyu Tan, Yicheng Wu, Ashok Veeraraghavan
3D cameras are rapidly gaining widespread adoption in a variety of applications. We propose a passive single-view 3D imaging system by inserting a phase mask at the aperture plane of a camera. We show an end-to-end optimization framework to joi
End-to-end Learned Phase Masks for Passive Single View Depth Estimation
Yicheng Wu, Vivek Boominathan, Huaijin Chen, Aswin C. Sankaranarayanan, Ashok Veeraraghavan
With the increasing demand for fast motion recognition, high-speed vision sensors are gaining widespread adoption. In particular, high-frame-rate three-dimensional (3D) imaging with additional depth information is rapidly becoming important in a variety of applications, such as automotive drive, virtual reality and video surveillance. One of the challenges of the high-frame-rate 3D cameras is the low image quality due to the limited amount of light reaching the sensor within a frame. Opening the aperture is a simple way to allow more light into the camera while inducing more out-of-focus blur as a trade-off. The objective of this project is to to develop a snapshot extended depth of field (EDOF) camera with a large aperture that enable the all-in-focus detection of a 3D scene, by inserting an optimized phase mask into the aperture plane of the camera. The phase mask will be learned to enforce depth-invariant point spread functions (PSFs) for all-in-focus reconstruction as well as encourage containing high frequency in PSFs for detailed feature recovery. In particular, the mask will be simulated by an end-to-end optical optimization framework which simultaneously optimize the mask height map and the reconstruction algorithms, and additional prior information will be adopted into the framework to enhance the performances.
Estimation of Human Circadian Phase Using Sensor Data with a Two-Step Framework
Cheng Wan, Akane Sano
The circadian rhythm is one of the most essential biological processes in the human body, and it affects various biological activities including skin temperature, hormone production, and nutrient absorption. The dim light melatonin onset (DLMO) is the gold standard for measuring the human circadian phase. The collection of DLMO is expensive and time-consuming as saliva or blood sampling is required during overnight studies in specialized laboratory settings. In the past few years, several non-invasive approaches have been designed for estimating DLMO values. These methods collect sensor data including light exposure, skin temperature, physical activity, and sleep patterns. And the researchers train machine learning models for estimating DLMO. However, previous studies only leverage either real-time physiological or behavioral data (data up to now) or sleep patterns (sleep data up to today). In this paper, we propose an innovative machine learning model for estimating DLMO using both of realtime data and previous sleep features with a two-step framework. The first step summarizes the data prior to the current day, while the second step combines this summary with human real-time data of the current day. We implement two models following this framework. Both models apply a recurrent neural network to the second step, while one of them utilizes an exponential moving average model for extracting features from sleep patterns, which is proposed in this paper, and the other one uses a recurrent neural network for summarizing the features of real-time data over the past several days. The experimental results show that our framework is significantly better than the models that use only human activity or sleep patterns. We have determined that our new method is superior because its root-mean-square error on test set shows significantly smaller errors.
FaceEngage: Robust Estimation of Gameplay Engagement from User-contributed (YouTube) Videos
Xu Chen, Li Niu, Ashok Veeraraghavan, Ashutosh Sabharwal
Measuring user engagement in interactive tasks can facilitate numerous applications toward optimizing user experience, ranging from eLearning to gaming. However, a significant challenge is the lack of non-contact engagement estimation methods that are robust in unconstrained environments. We present FaceEngage, a non-intrusive engagement estimator leveraging user facial recordings during actual gameplay in naturalistic conditions. Our contributions are three-fold. First, we show the potential of using front-facing videos as training data to build the engagement estimator. We compile FaceEngage Dataset with over 700 picture-in-picture, realisitic, and user-contributed YouTube gaming videos (i.e., with both full-screen game scenes and time-synchronized user facial recordings in subwindows). Second, we develop FaceEngage system, that captures relevant gamer facial features from front-facing recordings to infer task engagement. We implement two FaceEngage pipelines: an estimator trained on user facial motion features inspired by prior psychological works, and a deep learning-enabled estimator. Lastly, we conduct extensive experiments and conclude: (i)~certain user facial motion cues (e.g., blink rates, head movements) are engagement-indicative; (ii)~our deep learning-enabled FaceEngage pipeline can automatically extract more informative features, outperforming the facial motion feature-based pipeline; (iii)~FaceEngage is robust to various video lengths, users/game genres and interpretable. Despite the challenging nature of realistic videos, FaceEngage attains the accuracy of 83.8%.
Feasibility of Passive Eavesdropping in Massive MIMO: An Experimental Approach
Chia-Yi Yeh, Edward W. Knightly
Massive MIMO has the potential to thwart passive eavesdropping as the signals transmitted by a large antenna array become highly focused. Indeed, the impact of passive eavesdropping has been shown to be negligible when the number of base station (BS) antennas approaches infinity for independent Rayleigh channels. In this paper, we experimentally explore eavesdropping in Massive MIMO incorporating real-world factors including a limited BS antenna array size, potential correlation in over-the-air channels, and adaptation of modulating and coding schemes (MCS) over a discrete and finite set. Using a 96-antenna ArgosV2 BS, we (i) explore scaling the array size; (ii) identify eavesdropper advantages due to channel correlation and the resulting increase in array size required to mitigate this advantage; (iii) identify the “MCS saturation regime” as a vulnerability even with high SNR, (iv) characterize transmit power control counter strategies at the BS, and (v) explore the impact of a nomadic eavesdropper that moves to find the most favorable position.
Focusing Through Scattering Media
June Chen, Ashok Veeraraghavan, Ashu Sabharwal
While medical devices such as Optical Coherence Tomography (OCT) are widely in use today to look into human bodies, these technologies suffer from high scattering in tissue, which results in blurred images only a few scattering lengths beneath the skin surface. Thus, it is vital to intelligently create a sharp, interpretable result. Common methods to solve this problem include directly inverting the blurred output based on the scattering media and adaptively focus the incoming beam to focus in the tissue. Since the exact composition of the tissue changes over time and remains unknown, it is hard to reconstruct the corrupted image based on the medium. On the other hand, adaptive focusing operates by shaping the input light based on the output directly, but the current method require many iterations and is slow. My research objective is to design an efficient adaptive focusing algorithm based on optimization methods and to build the hardware system from the computer simulation to demonstrate its functionality. The optimization algorithm faces several hardware constraints. For example, the phase of the output cannot be measured. Through an iterative process, the output would become in focus with the help of a spatial light modulator. With a correctly designed system, the efficiency of adaptive focusing would be increased by at least ten folds, and would increase diagnosis accuracy while drastically improving patient experience.
HodgeNet: Graph Convolutional Networks in the Edge Space
Mitch Roddenberry, Santiago Segarra
In an increasingly connected world, graphs have emerged as a powerful tool for modeling social interactions, disease propagation, and a variety of signals with relational structure. Recently, convolutional neural networks have been generalized to process data on graphs, with cutting-edge results in traditional tasks such as node classification and link prediction. However, these methods have all been formulated in a way suited only to data on the nodes of a graph, based on existing work in spectral graph theory. Using tools from algebraic topology, it is possible to reason about oriented data on higher-order structures, such as the edges or faces of a graph. The research objective is to develop techniques for applying the Hodge Laplacian to processing data on higher-order graph structures using convolutional neural networks. To test the efficacy of these techniques, the problem of flow interpolation will be tackled: that is, given observations of flow over a subset of the edges of a graph, how can flow over the unobserved edges be inferred? This has applications in a variety of contexts, such as in the analysis of traffic flow, economic flow, or the flow of resources in ecological systems.
Hybrid Photonic-Plasmonic Resonator System for Non-Hermitian Selective Thermal Emission
Chloe F. Doiron, Alex Hwang, and Gururaj V. Naik
Hermitian systems possess real eigenvalues and hence describe real world systems. However, Hermiticity is only a sufficient condition for a real eigenvalue spectrum. Non-Hermitian systems with parity-time (PT) symmetry have been shown to exhibit real eigenvalues. Such non-Hermitian systems possess exceptional points in their phase space and exhibit interesting topological properties. Selective thermal emitters emit thermal radiation in a narrow frequency range, making them an exciting light source for sensing and waste heat recovery through thermophotovoltaics. Since selective thermal emitters exchange energy with their environment, through thermal radiation, they are always non-Hermitian making them an ideal platform to study non-Hermitian physics. While selective thermal emitters are always non-Hermitian, there have been no experimental demonstrations of explicitly non-Hermitian behavior. In this work, we experimentally demonstrate passive PT-symmetry in selective thermal emission at 700°C by strong coupling photonic and plasmonic resonators. Passive PT-symmetry occurs in the coupled system due to the extreme asymmetry in optical losses. By tuning the coupling between the resonators, we demonstrate the ability to tune the system between the PT-symmetric and broken PT-symmetry phases. Furthermore, by controlling the internal oscillator phase we were able to control far-field thermal radiation. The ability to tune far-field thermal emission by controlling the internal oscillator phase opens new pathways for engineering high performance selective thermal emitters for sensing and waste heat recovery applications.
Image Representation Capacity of Untrained Non-Convolutional Neural Networks
Paul Mayer, Reinhard Heckel, Rich Barabiuk
Studying the image representation capacity of untrained non convolutional neural networks.
In-Memory Computing Architecture for Deep Neural Networks
Zhiyu Chen, Kaiyuan Yang
To achieve higher energy efficiency, state-of-the-art in-memory computing DNN accelerators modify the conventional Von Neumann architecture by implementing Multiply-Accumulate (MAC) operations in analog domain inside the memory, and therefore avoid the energy-hungry data movement. However, these architectures suffer from either restricted choice of neural networks, which only support naive models such as binarized neural networks, or relatively low accuracy. Meanwhile, a family of algorithms called model compression demonstrate great potential to simplify complicated DNN structure in software community, but current GPU/CPU is unable to implement the algorithms in an optimized way, thus they are in urgent need of customized hardware. Our research goal is to design a reconfigurable and accurate in-memory computing architecture using SRAM that supports corresponding co-designed model compression algorithms. The proposed reconfigurable architecture supports in-memory MAC operations of multi-bit input feature map and multi-bit weights based on the structure of the compressed model, while maintaining high accuracy and energy efficiency.
Joint MIMO Communications and Wireless Imaging
Nate Raymondi, Cheng Li, Ashutosh Sabharwal
In this paper, we analyze the performance of a joint MIMO node capable of simultaneous radar and communication functions. We derive outer bounds for rate regions of communication and radar estimation in terms of the communication link’s information rate and the formulated radar estimation rate – a measure of how much information can be extracted from the radar return. We show that the knowledge of the communication signal can be utilized to improve the radar estimation rate. A legacy system’s performance suffers as a radar target moves closer to a communications user due to increased interference from beamforming leakage power. However, we show that a joint system can benefit from this extra power. Since the joint system perfectly knows the radar and the communication signals, radar performance can be enhanced using communications returns from the radar target.
Leaky Waveguide Arrays: Rotational Resilience in Leaky Waveguide based THz WLANs
Mohammad Furqan Ahmad
Parallel plate leaky waveguide (LWG) antennas operating in the 0.1-1.0 Terahertz (THz) spectrum have attracted growing interest in recent years as promising technology for enabling Tb/second wireless communication networks. LWGs are rectangular waveguides with a slot on one face, through which the radiation from the waveguide can ‘leak’ into free space. The angle at which a signal is coupled from the waveguide into free space (or vice versa) is directly linked to the carrier frequency and bandwidth of the signal. When operated in the THz spectrum, this angle-frequency coupling in LWGs leads to highly directional beams, which are required to counter the high levels of signal attenuation in the THz band. In the context of employing LWGs in THz WLANs, recent research has focused on experimentally achieving Tb/second data rates in fixed wireless links where both the Access Point (AP) and the client employ LWGs and are aligned i.e. the angle of departure of the THz beam from the AP-side LWG matches the angle of arrival of the beam at the client’s LWG. However, if the AP and client are not aligned, there is a frequency-angle mismatch, leading to a fall in achievable physical layer rate. To mitigate this effect, I propose employing an array of switchable LWG antennas at the AP. My research objective is to simulate how the average achievable physical layer rate varies as a function of the number of LWG antennas employed at the AP in a point-to-point LWG-based THz wireless link, first employing only line-of-sight paths and then incorporating non-line-of-sight paths in the simulation scenario.
Light Phi Phenomenon and Subject Intervention
Luis Hector Victor
Alertness, also known as sustained attention, is a critical component of cognition. It is the ability to maintain selective and focused attention and response with time under potentially variable conditions. This cognitive ability is far-reaching, from playing a vital role in surgery rooms to military surveillance, and thus has been thoroughly studied for optimization. Static, dynamic, or temporally modulated, and specific wavelengths, colors, of light have been shown to improve alertness. However, studies indicate that there is still a demand for improved intervention methods for preventing alertness decrement. Then, here we argue for a light phi phenomenon intervention method for alertness maintenance and improvement. Phi phenomenon, similar to beta movement, describes the perception of continuous motion due to temporally, spatially variant images. We will investigate three models: apparent diffusive light phi phenomenon, directed light phi phenomenon, and random light patterns to understand how light phi phenomenon affects a subject’s alertness and how it compares to traditional light intervention methods.
Magnetoelectric Materials For Miniature, Wireless Neural Stimulation At Therapeutic Frequencies
Amanda Wickens, Benjamin Avants, Nishant Verma, Eric Lewis, Joshua C. Chen, Ariel K. Feldman, Shayok Dutta,
Joshua Chu, John O’Malley, Michael Beierlein, Caleb Kemere, Jacob T. Robinson
A fundamental challenge for bioelectronics is to deliver power to miniature devices inside the body. Wires are common failure points and limit device placement. Wireless power by electromagnetic or ultrasound waves must overcome absorption by the body and impedance mismatches between air, bone, and tissue. Magnetic fields, on the other hand, suffer little absorption by the body or differences in impedance at interfaces between air, bone, and tissue. These advantages have led to magnetically powered stimulators based on induction or magnetothermal effects. However, fundamental limitations in these power transfer technologies have prevented miniature magnetically-powered stimulators from applications in many therapies and disease models because they do not operate in clinical high frequency ranges above 50 Hz. Here we show that magnetoelectric materials, applied for the first time in bioelectronics devices, enable miniature magnetically-powered neural stimulators that operate at clinically relevant high-frequencies. As an example, we show that ME neural stimulators can effectively treat the symptoms of a Parkinson’s disease model in a freely behaving rodent. We also show that ME-powered devices can be miniaturized to sizes smaller than a grain of rice while maintaining effective stimulation voltages. These results suggest that ME materials are an excellent candidate for wireless power delivery that will enable miniature neural stimulators in both clinical and research applications.
Microfluidic actuation of flexible microelectrodes for neural recording
Bo Fan, Alexander V. Rodriguez, Daniel Vercosa, Caleb Kemere, and Jacob Robinson
Small, flexible neural electrodes with diameters the size of individual cells significantly increase the quality and longevity of neural recordings by reducing neural injury during chronic implantation; however, flexible electrodes are traditionally difficult to implant and position without causing acute damage. Implantation of flexible electrodes typically requires stiffening agents that temporarily increase the overall size and rigidity of the electrode in order to overcome the buckling force upon implantation. The resulting increased electrode footprint leads to cell loss and glial activation that persists even after the stiffening agents are removed or dissolved. Recently, we have developed a device for minimally invasive implantation of multi-channel flexible electrodes using a specially designed microdrive utilizing microfluidics that eliminates the need for stiffening agents or shuttles. In the fluidic microdrive, microfluidic channels act as mechanical constraints preventing buckling while viscous drive fluid pushes the electrode into the brain by exerting drag force along the length of the probe. Vent channels redirect and significantly decrease the amount of fluid that exits towards the brain, and slow fluid speeds due to viscosity prevent drive fluid from entering or damaging the brain. Our current work focuses on quantifying and comparing acute damage and immune response following in vivo implantation of flexible electrodes using traditional stiffener-assisted methods and our fluidic microdrive.
Miniaturized lens-less microscope for optical neural signal stimulating and recording
Dong Yan, Jesse K. Adams, Jacob T. Robinson
The aim of my project would be to construct a miniaturized lens-less microscope consisting of both image-capturing hardware and image-processing software. The construction and optimization of this microscope would be focused on optical neural signal stimulating and recording. The miniaturization feature of such microscope would allow in vivo imaging of neural activity in small animals, e.g. rats, with minimized constraint to their movement, while the lens-less imaging technology would allow optical signals in a 3D volume to be reconstructed from a raw image taken with a single shot. This work would realize real-time detection of optical neural signals in a 3D volume.
Multi-parameter Optimization of Neural Activity Biosensors by a High-throughput Screening Platform
Zhuohe Liu, Yueyang Gou, Sihui Guan, Xiaoyu Lu, Jihwan Lee, Andreas S. Tolias, Francois St-Pierre
This work provides a fast and effective way to optimize biosensors of neural activity, which is critical for understanding functions of neuron circuits. Engineered voltage-sensitive fluorescent proteins, termed genetically-encoded voltage indicators (GEVIs), have the capability to enable voltage recordings in vivo with subcellular resolution and cell type specificity. However, GEVIs are not optimal for long-term recording due to low brightness and lack of photostability, and their sensitivity and kinetics are not sufficient for detection of fast voltage dynamics in vivo. This work presents an improved indicator, JEDI-Dα, discovered by an automatic high-throughput screening platform. The platform scores multiple performance metrics for mutagenesis libraries in a 96-well format, achieving a 100-fold speed-up over traditional laborious electrophysiological methods. Further experiments demonstrated that the brighter, faster, more sensitive, and two-photon compatible JEDI-Dα enables accurate detection of neural electrical activity in dissociated neurons and organotypic slice cultures with sub-millisecond temporal resolution and sub-micron spatial detail.
Networked Drones System for Air Monitoring
Maryam Khalid, Edward W. Knightly
Uncontrolled emissions of gases in industries from accidents and disasters result in loss of life and property. Even if the situation is not critical, everyday unchecked emissions still affect the quality of air and result in air pollution. In both these scenarios, a reliable survey of toxic substances with high spatial and temporal resolution is required. Three candidate technologies that have been widely studied for gas sensing applications include wireless sensor networks (WSN), mobile robots and Unmanned aerial vehicles (UAV). WSN can not cater to the dynamic nature of gas transport process which is to some extent addressed by mobile robots. However, both of them are restricted to 2-D exploration. Therefore, UAVs are the most suitable option because they are light-weight, easy to deploy and allow coverage in 3-D with accessibility to difficult-to-reach places. Although single-UAV based systems have been explored in literature for gas sensing applications, research on deployment of a UAV network, which is more robust and fault-tolerant, for such applications is still in infancy. This project aims to provide an end-to-end solution for characterization of gaseous plumes through optimal deployment of UAV network. To guarantee coverage of the complete region, a vehicle routing problem is formulated. The concentration measurements from the initial flight of UAVs are used to construct the gas distribution map using Gaussian kernel extrapolation. Further, the path is planned adaptively to refine this coarse estimate. Finally, the true and estimated distribution maps are compared to evaluate the performance of the system.
Phase calibration in bistatic mmWave wireless imaging
Tianyu Cao, Ashutosh Sabharwal
Wireless signals always contain a lot of information that can be used to do communications and many other works together. Imaging is one of these works that has shown prospect in many applications. Using tools like mmWave band and phase-based Fourier optics are good ways to get images with good resolution without reducing network performance. Previously, high resolution could be achieved in monostatic mmWave wireless imaging system [Sheen et al], where transmitters and receivers are placed together. However, this method cannot be directly used in bi-static scene, where transmitters and receivers are placed separately, because of the existence of extra phase difference like unknown transmitting messages, unknown transmitter position, different oscillators at transmitters and receivers and so on. The objective of this research is to develop methods to calibrate these extra phase differences in bi-static mmWave wireless imaging systems. Furthermore, methods that use the calibrated phase information to do imaging in wireless networks will be developed. Simulations will be performed to show that we can both get high resolution images and maintain network performance through our wireless imaging methods.
Physically Based Rendering for Simulating Differential Images of Dynamics in Biological Tissue
Yongyi Zhao, Adithya Pediredla, Ashok Veeraraghavan
There is a need to improve the efficiency of simulating a differential image: i.e. determining a time-varying signal from the difference between multiple frames of a dynamic event. An efficient simulation would significantly improve the development of novel biomedical devices, such as high-resolution functional imaging. Current methods for such simulation often use standard Monte Carlo (sMC), which offers high accuracy and flexibility, but requires a long runtime. We propose to use techniques from physically based rendering, such as next event estimation (NEE), for the novel application of rendering a differential image. We specifically deploy the application for simulation of photon propagation in biological tissue and show that our technique can produce orders of magnitude speed-ups over sMC.
Quantitative Description of Nanorod Aggregates in Scanning Electron Microscopy Images
Rashad Baiyasi, Miranda J. Gallagher, Qingfeng Zhang, Stephan Link, Christy F. Landes
Aggregation is a major concern when working with colloidal suspensions of nanoparticles. Despite extensive research into the conditions in which nanoparticle aggregation occurs, little has been reported on the inter-particle structure of the aggregates that do form. We have developed two methods for quantitatively measuring the physical structure of nanoparticle aggregates: an algorithm for segmenting dense aggregates measured with scanning electron microscopy (SEM) and an order parameter for characterizing the side-by-side structure. The segmentation algorithm is an application of the marker-controlled watershed method where the nanoparticle markers are isolated through a series of image-processing steps. We have successfully segmented individual nanoparticles in aggregates under conditions with dim boundaries and intensity variation that preclude the use of other methods. Segmented SEM images can be used to quickly calculate the side-by-side order of a large number of aggregates. We report on the differences in gold nanorod side-by-side order after induced aggregation with bovine serum albumin and salt (NaCl) Future work will see these methods implemented with an open-source, user-friendly interface to provide quantitative image processing tools for researchers to characterize aggregate structure with high throughput.
RACE: Sub-linear memory sketches for approximate near-neighbor search on streaming data
Benjamin Coleman, Anshumali Shrivastava, Richard G. Baraniuk
We demonstrate the first possibility of a sub-linear memory sketch for solving the approximate near-neighbor search problem. In particular, we develop an online sketching algorithm that can compress N vectors into a tiny sketch consisting of small arrays of counters whose size scales as O(N^b log^2 N), where b<1 depending on the stability of the near-neighbor search. This sketch is sufficient to identify the top-v near-neighbors with high probability. To the best of our knowledge, this is the first near-neighbor search algorithm that breaks the linear memory (O(N)) barrier. We achieve sub-linear memory by combining advances in locality sensitive hashing (LSH) based estimation, especially the recently-published ACE algorithm, with compressed sensing and heavy hitter techniques. We provide strong theoretical guarantees; in particular, our analysis sheds new light on the memory-accuracy tradeoff in the near-neighbor search setting and the role of sparsity in compressed sensing, which could be of independent interest. We rigorously evaluate our framework, which we call RACE (Repeated ACE) data structures on a friend recommendation task on the Google plus graph with more than 100,000 high-dimensional vectors. RACE provides compression that is orders of magnitude better than the random projection based alternative, which is unsurprising given the theoretical advantage. We anticipate that RACE will enable both new theoretical perspectives on near-neighbor search and new methodologies for applications like high-speed data mining, internet-of-things (IoT), and beyond.
Rank aggregation from pairwise comparison
Gaurav Gupta, Anshumali Shrivastav
For many applications like ranking online players, aggregating social opinions and recommendation systems the objects are usually labeled with mutual preferences and comparisons. Ranking algorithms use these pairwise comparisons to form a global ranking of objects. The challenge is to rank the objects with fewer comparisons as possible. My project targets pair selection based on the similarity and dissimilarity between the object representation in a n dimensional feature space. The existing and most successful methods frame the approach assuming Bradley-Terry- Luce model of the system, where each object is associated with a score ranges from 0 to 1. The score represents the probability of an object being preferred over any other given object. The state of art methods for active ranking uses lower upper confidence bounds for finding top k objects. Some of the methods also modeled the problem as a graph where each vertex is an object and the pairwise comparison is a node, and the goal is to achieve a sparse graph. I have used locality sensitive hashing (LSH) for sampling pairs in the rank aggregation and hence implement a fast and approximate ranking algorithm.
Recovering Data in a DNA Data Storage System
Mohammad Z. Darestani, Reinhard Heckel
Employing Min-Hashing method as an error-correcting algorithm to increase the accuracy of recovering the original data from the DNA sequences existing in the storage system.
Secure Process Abstract
Caihua Li, Lin Zhong
This project explores the ARM TrustZone technology to rethink the systems software so that application secrets are shielded from the underlying operating system. Operating systems are complex software with a wide attack surface; security vulnerabilities of widely used OSes are regularly discovered and exploited. Moreover, as many applications can be purchased freely, an attacker can simply run them in a modified operating system to steal their secrets. The problem with today’s systems software is that the OS is in charge of both protection and resource management. This project seeks to decouple these two functions and move part of protection out of OS and into a trusted piece of software, i.e., an application-agnostic program running in the Secure World. We have demonstrated some limited success of this approach in Ginseng. However, Ginseng suffers from the following problems: (1) it features an awkward programming model in which the compiler and the developer collaborate to identify sensitive data and keep them in the registers. (2) its runtime protection of sensitive data in the registers incur considerable overhead. In this project, we aim at investigating a secure process abstraction. When an application is launched as a secure process, its entire virtual address space will be shielded from the OS. Like Ginseng, we rely on an application-agnostic, trusted program in the Secure world to provide protection. The Normal world OS only performs resource management for the secure process. We will start with a straightforward extension of Ginseng, which is expected to have extremely high overhead due to memory access and world switching. We will then focus on optimizing this baseline design with both systems and architecture/hardware innovations.
Single-Frame Spatio-temporally Resolved Reconstruction of Single Particle Trajectories
Jorge Zepeda O, Christy Landes
To advance our understanding of molecular dynamics in both chemical and biological systems, we require the ability to localize individual particles in both three-dimensional space and time beyond the diffraction limit and camera frame rates using wide-field microscopic imaging. Many methods such as PALM or STORM provide nanoscale spatial resolution, but sacrifice temporal resolution.The challenge is to obtain a four-dimensional image of each particle despite planar projection and temporal integration. A solution is to design a phase mask that duplicates the point spread function, forming a two-lobed image for one particle. Prior work (Wang, et al., 2016, Landes, 2017) was only able to encode either the depth or temporal component by the lobes’ rotation. However, a new phase mask considers both inter-lobe distance and rotation, allowing for information in all four dimensions to be encoded. My research is to generalize the current three-dimensional approach to resolve both depth and temporal information from a single frame and form a trajectory. I will first resolve the image into its space-time coordinates using an L1-based deconvolution algorithm, obtaining a sparse list of points. These points will be sorted by time and de-noised using a local confidence metric. Then the data is sliced into temporal hyperplanes and fit each slice with an arbitrarily transformed parabolausing linear least squares regression. After determining these fits, I can interpolate the overall space-time trajectory of the particle. This approach will be tested first in simulation with 3D and 4D trajectories of varying complexity in order to establish its accuracy and limitations. Afterwards, I will verify the results experimentally by fixing polystyrene beads and maneuvering them around using a piezoelectric stage.
Social Acoustic Ambiance Understanding
Wenwan Chen, Jian Cao, Ashutosh Sabharwal
Since social environment influences the manifestation of personality in sociability, it is of great significance to measure social ambiance via modern, objective, scalable methods. Audio recordings can serve as a potential entry point, with the help of deep learning. The challenge is to connect the properties of the audio data to the social behavior knowledge base. Specifically, I will capture features like sound duration, harmonicity as well as high-level VGG embedding features, to interpret audio data in a socially meaningful way. Model will be built and trained using out-of-domain dataset, and then transferred to our own dataset. Finally it will be shown that the proposed method can examine the relation between sociability of three groups across the lifespan and psychopathology, which will serve as a tool to assist clinicians in the decision making process.
SparsePPG: Towards Driver Monitoring Using Camera-Based Vital Signs Estimation in Near-Infrared
Ewa Magdalena Nowara, Tim K. Marks, Hassan Mansour, Ashok Veeraraghavan
Camera-based measurement of the heartbeat signal from minute changes in the appearance of a person’s skin is known as remote photoplethysmography (rPPG). Methods for rPPG have improved considerably in recent years, making possible its integration into applications such as telemedicine. Driver monitoring using in-car cameras is another potential application of this emerging technology. Unfortunately, there are several challenges unique to the driver monitoring context that must be overcome. First, there are drastic illumination changes on the driver’s face, both during the day (as sun filters in and out of overhead trees, etc.) and at night (from streetlamps and oncoming headlights), which current rPPG algorithms cannot account for. We argue that these variations are significantly reduced by narrow-bandwidth near-infrared (NIR) active illumination at 940 nm, with matching bandpass filter on the camera. Second, the amount of motion during driving is significant. We perform a preliminary analysis of the motion magnitude and argue that any in-car solution must provide better robustness to motion artifacts. Third, low signal-to-noise-ratio (SNR) and false peaks due to motion have the potential to confound the rPPG signal. To address these challenges, we develop a novel rPPG signal tracking and denoising algorithm (sparsePPG) based on Robust Principal Components Analysis and sparse frequency spectrum estimation. We release a new dataset of face videos collected simultaneously in RGB and NIR.We demonstrate that in each of these frequency ranges, our new method performs as well as or better than current state-of-the-art rPPG algorithms. Overall, our preliminary study indicates that while driver vital signs monitoring using cameras is promising, much work needs to be done in terms of improving robustness to motion artifacts before it becomes practical.
Structured Illumination for Low Scattering Imaging in Low-cost System
Mary Jin, Yubo Tang, Ashok Veeraraghavan
We designed and experimented with a low-cost microscopy system that can provide cellular resolution images with high contrast by using colored structured illumination. The system can image in real-time, which could potentially be used in vivo for cancer screening.
Thermal Near –field Manipulation Using Far-field Optical Excitation
Seyeyd Ali Hosseini Jebeli,Ujjal Batacherjee, Wei-Shun Chang, and Stephan Link
Plasmonic nano structures are used in various applications such as SERS, SEIRA and chemical sensors due to their strong electric field localization and enhancement in their near-field region. The restricting factor for field enhancement is the loss due to damping of the plasmon just a few tens of femtoseconds after its excitation. Recently, the hot carriers generated after plasmon damping were used for photocatalysis and it was a new approach to take advantage of the loss in plasmonic nanostructures. But most of the plasmon energy is converted to heat and is dissipated through the surrounding medium thus finding a way to use the generated heat will lead to new applications. In this work, it is demonstrated using photothermal microscopy that by changing wavelength and polarization of the incident light, the thermal near-field of the plasmonic nano structures can be controlled and significant temperature gradient is created over gold nano rod heterodimers. These new findings provide a new platform to do temperature sensitive chemical reactions at nano scale and pave the way toward nano scale material engineering using visible and infra-red light.
Towards a unified theory of information processing in resource-constrained brain circuits
Lokesh Boominathan, Xaq Pitkow
The purpose of this project is to explain the brain quantitatively, by appealing to the principles of brain function. There are mainly four theories that were proposed in this direction – predictive coding, efficient coding, message passing, and inference. Although these theories were applied in different contexts, we hypothesize that we can unite them by formulating a simple model where everything should be analytically solvable and should give us the ability to test the core foundational ideas. For this, we would be using probabilistic graphical models, with power constraints, and some basic assumptions such as Gaussian and linearity to make it tractable. This would have consequences on modeling how neurons respond to novel inputs and thereby lead to a good theory of brain function with potential applications in both medicine and engineering. For example, understanding how neurons process information under resource-constrained settings can be used for building smart devices with low power requirements.
Towards understanding the neural mechanisms of haptic communication
Alix Macklin, Marcia O’Malley
Communication via haptic (tactile) channels provide a mode of communication for the visually and/or hearing impaired and may offer alternative pathways for language rehabilitation and development in individuals with neurological impairments as well as language processing or developmental disorders. Due to developments in wearable technology, progress in haptics research, and increased understanding of tactile perception due to advanced neuro-imaging and recording techniques, there is recent interest in understanding how the brain processes language via haptic language devices. The Mechatronics And Haptics Interfaces (MAHI) lab at Rice University has recently developed the MISSIVE (Multi-sensory Interface of Stretch, Squeeze, and Integrated Vibration Elements), a wearable, haptic device that transmits English phonemes to the user via multi-feature cues. Unlike existing tactile-communication devices, the MISSIVE renders cues that incorporate lateral skin stretch and radial squeeze, as well as typical vibrotactile actuation. MISSIVE generates haptic cues that were found to improve perceptual distinguishability among users compared to cues generated by vibrotactile components alone. Despite the promise of this device, our knowledge on how individuals process language via haptic actuation is confined to solely vibrotactile encodings of speech. To successfully integrate haptic language devices in assistive and rehabilitative technologies, it is imperative we understand how tactile language is processed in the brain using multi-feature devices as well. First, we aim to understand how perception to multi-feature cues changes with training. Second, we aim to understand how multi-feature haptic cues are mapped to language. Using MISSIVE as the mode of tactile communication and Electroencephalography (EEG) to record neural activity to unique MISSIVE cues, we will rigorously characterize neural responses during multi-feature, tactile language transmission. First, we will assess the Mismatch negativity (MMN) response to determine how the neural response to haptic cues change after such cues have been mapped to language-specific phonemes, via a validated phonetic training regimen. We will use an oddball paradigm to compare the MMN responses to unique multi-feature, haptic cues before phoneme training and after training. An overall increase in MMN response amplitude after phoneme training would suggest the haptic cue set is encoded similarly to auditory language. Second, we will determine the lexicality effect of phoneme-based word training on the neural responses to unique haptic cues. We expect unique memory traces to be elicited for words that are conveyed haptically versus non-words. Specifically, we will compare the MMN responses between haptic phoneme-cues that form words to those that form non-words (e.g. ‘car’, ‘cat’ versus ‘cas’, ‘cak’), and we expect to see enhanced MMN responses when the final phoneme cue completes a known English word, suggesting haptic language is processed similarly to auditory speech.
Virtual Speed Test: an AP Tool for Passive Analysis of Wireless LANs
Peshal Nayak, Santosh Pandey, Edward W. Knightly
Internet speed tests assess end-to-end network performance by measuring throughput for 10s of MB of TCP uploads and downloads. While such tests provide valuable insights into network health, they are of little use to network
administrators since (1) the results are only available on the client that performs the test and (2) the tests can saturate the network, increasing load and worsening performance for other clients. In this paper, we present virtual speed test, a measurement based framework that enables an AP to estimate speed test results for any of its associated clients without any special-purpose probing, with zero end-user co-operation and purely based on passively observable parameters at the AP. We implemented virtual speed test using commodity hardware, deployed it in office and residential environments, and conducted
measurements spanning multiple days having different network loads and channel conditions. Overall, virtual speed test has mean estimation error less than 6% compared to ground truth speed tests, yet with zero overhead, and outcomes available at the AP.
Wellbeing Prediction using Sensors/Phone data with Machine Learning
Han Yu, Akane Sano
We built and compared several machine learning models to predict future self-reported wellbeing labels (of mood, health, and stress) for next day and for up to 7 days in the future, using multi-modal data. The data are from surveys, wearables, mobile phones and weather information collected in a study from college students, each providing daily data for 30 or 90 days. We compared the performance of multiple models, including personalized multi-task models and deep learning models. The best personalized multi-task linear model showed mean absolute errors of 12.8, 11.9, and 13.7 on a continuous-100 pt scale for estimating next day’s mood, health, and stress value, while the best multi-task neural network model, applied to 3-way high/med/low classification of the wellbeing values showed F1 scores of 0.71, 0.74, and 0.66 on mood, health, and stress metrics, respectively.
Zero Shot Learning with Joint Visual-Semantic Embedding
Shabnam Daghaghi, Anshumali Shrivastava
Todays Neural Networks, specifically CNNs, play a crucial role in image classification and they are known to require numerous training samples to have promising performance. Few/zero shot learning problems arise when few/no samples are available for specific classes e.g. when a new product is added to Amazon. Typically zero shot datasets include semantic representations/attributes in addition to images/features and labels. Current state-of-the-art approaches include embedding space transformation/alignment (independently finding embeddings of visual and semantic data then either transforming one to another or aligning the two), and data augmentation usually with Generative Adversarial Networks (generating synthetic samples for the unseen/sparse class via either generating the image itself or the visual features). However the current methods neither are optimal in the sense of joint embedding space nor scalable for extremely large data sets. In my research I am going to find the optimal joint embedding space and also develop an algorithm to make the proposed method scalable. In order to achieve this objective I will set up an optimization framework with appropriate loss function to simultaneously obtain a joint embedding of multimodal data (e.g. visual and semantic) and learn the classifier. Additionally I am going to employ randomized hashing and sketching algorithms to scale the algorithm to enormously large data sets and fine- grained classification regime.
POSTDOCS/GUESTS/STARTUPS
Automated Sports Analytics
James Grinage, Connor Heggie
In 2002, Moneyball and Sabermetrics thrust the sports world into the age of analytics. Unfortunately, access to these analytics is costly, time-consuming and inaccurate. We at Cherrypick believe that with faster and more accurate game statistics, coaches, players, and teams can maximize their winning potential. That’s why we’ve developed state-of-the-art machine learning and computer vision techniques to automate full game breakdowns to provide accurate, same-day sports statistics.
Numina: Real-time insights from streets
Ilan Goodman, Jennifer Ding
Numina’s mission is to make cities more responsive, so they are safer, healthier, and more equitable for the people who live in them. We make a hardware and software computer vision platform that senses bicycle, pedestrian, and other street traffic and delivers real-time intelligence — without surveillance — to help urban planners and municipal governments design better streets and public places. Join our small but quickly growing team, as we further design, develop, and deploy a system that has already measured 30 million trips since October 2018.
RENEW Programmable and Observable Massive-MIMO platform
Rahman Doost-Mohammady, Oscar, Bejarano, Clayton Shepard, Ryan Guerra, Lin Zhong, Ashutosh Sabharwal
The RENEW project will develop world’s first fully programmable and observable wireless radio network. With RENEW, wireless research and development community will be able to test diverse ideas and concepts, ranging from low-level hardware to all the way to novel applications. RENEW uses ArgosV3 Massive-MIMO base station hardware from Skylark Wireless (a startup spin-off of Rice) and develops a diverse set of software stacks for wireless research. In this demo, we showcase multiple frameworks developed by the RENEW team as useful research tools for Massive-MIMO research, including MATLAB and Python frameworks for non-realtime massive-MIMO research as well as Real-time framework doing beamforming to multiple users as a time.
TIME: Pushing Most Data Movements and Interfaces in ReRAM-based PIM Accelerators to be Local and Analog
Weitao Li, Yingyan Lin
Resistive-random-access-memory (ReRAM) based processing-in-memory (PIM) accelerators have emerged as a promising solution to bridge the gap between the constrained resources of Internet of Thing (IoT) devices and the prohibitive energy cost of deep neural networks (DNNs). Specifically, ReRAM-based DNN accelerators enhance the 1) energy efficiency by removing the data movement cost of weights through PIM and 2) computation density thanks to the high density of ReRAM, as compared to digital DNN accelerators. However, the large energy cost associated with 1) data movements of inputs and partial-sums and 2) interfacing circuits of ReRAM-based DNN accelerators still largely limit their achievable energy efficiency. To this end, we propose an ReRAM-based PIM DNN architecture, TIME, that can aggressively reduce the necessity of high-cost data movements of inputs and partial-sums and exploits more efficient interfacing circuits. Analytical models are proposed for evaluating ReRAM-based PIM DNN accelerators in terms of energy efficiency, throughput, and area. Compared with state-of-the-art ReRAM-based accelerators, TIME improves a peak computation density by at least 14.6x, and achieves at least 9.0x higher energy efficiency and 38.9x higher throughput, when evaluated in various CNN models and benchmark datasets.
Wellbeing Prediction using Sensors/Phone data with Machine Learning
Han Yu; Akane Sano
We built and compared several machine learning models to predict future self-reported wellbeing labels (of mood, health, and stress) for next day and for up to 7 days in the future, using multi-modal data. The data are from surveys, wearables, mobile phones and weather information collected in a study from college students, each providing daily data for 30 or 90 days. We compared the performance of multiple models, including personalized multi-task models and deep learning models. The best personalized multi-task linear model showed mean absolute errors of 12.8, 11.9, and 13.7 on a continuous-100 pt scale for estimating next day’s mood, health, and stress value, while the best multi-task neural network model, applied to 3-way high/med/low classification of the wellbeing values showed F1 scores of 0.71, 0.74, and 0.66 on mood, health, and stress metrics, respectively.