BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Department of Electrical & Computer Engineering - ECPv6.15.20//NONSGML v1.0//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
X-WR-CALNAME:Department of Electrical & Computer Engineering
X-ORIGINAL-URL:https://ece.northeastern.edu
X-WR-CALDESC:Events for Department of Electrical & Computer Engineering
REFRESH-INTERVAL;VALUE=DURATION:PT1H
X-Robots-Tag:noindex
X-PUBLISHED-TTL:PT1H
BEGIN:VTIMEZONE
TZID:America/New_York
BEGIN:DAYLIGHT
TZOFFSETFROM:-0500
TZOFFSETTO:-0400
TZNAME:EDT
DTSTART:20200308T070000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0400
TZOFFSETTO:-0500
TZNAME:EST
DTSTART:20201101T060000
END:STANDARD
BEGIN:DAYLIGHT
TZOFFSETFROM:-0500
TZOFFSETTO:-0400
TZNAME:EDT
DTSTART:20210314T070000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0400
TZOFFSETTO:-0500
TZNAME:EST
DTSTART:20211107T060000
END:STANDARD
BEGIN:DAYLIGHT
TZOFFSETFROM:-0500
TZOFFSETTO:-0400
TZNAME:EDT
DTSTART:20220313T070000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0400
TZOFFSETTO:-0500
TZNAME:EST
DTSTART:20221106T060000
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210910T140000
DTEND;TZID=America/New_York:20210910T150000
DTSTAMP:20260426T130255
CREATED:20210908T194338Z
LAST-MODIFIED:20210908T194338Z
UID:5156-1631282400-1631286000@ece.northeastern.edu
SUMMARY:ECE PhD Dissertation Defense: Wenqian Liu
DESCRIPTION:PhD Dissertation Defense: Explainable Efficient Models for Computer Vision Applications \nWenqian Liu \nLocation: Zoom Link \nAbstract: State of the art deep learning based models\, such as Convolutional Neural Networks (CNNs) and generative models\, achieve impressive results\, but with their great performance comes great complexity and opacity\, huge parametric spaces and little explainability. The criticality of model explainability and output interpretability\, manifests clearly in real-time critical decision making processes and human-centred applications\, such as in healthcare\, security and insurance. Explainability and interpretability are tackled in this thesis\, as intrinsic qualities in the model architecture as well as post-hoc improvement on existing models. In the area of frame prediction in video sequences\, we introduce DYAN\, a novel network with very few parameters\, that is easy to train and produces accurate high quality predictions. Another key aspect of DYAN is interpretability\, as its encoder-decoder architecture is designed following concepts from systems identification theory and exploits the dynamics-based invariants of the data. We also introduce KW-DYAN\, an extension of DYAN that tackles the issue of time lagging in video predictions\, by implementing a novel way of quantifying prediction timeliness and proposing a new recurrent network for adaptive temporal sequence prediction. The experimental results show the reduced lagging across datasets\, while also performing well in other metrics. In this thesis we also propose the first technique to visually explain VAEs by means of gradient-based attentions\, with methods to generate visual attentions from the learned latent space\, and also demonstrate such attention explanations serve more than just explaining VAEs. We show how these attention maps can be used to localize anomalies in images\, conducting state-of-the-art performance on multiple datasets. We also apply our technique for skin image anomaly detection and diagnosis and achieve competitive quantitative and qualitative results.
URL:https://ece.northeastern.edu/event/ece-phd-dissertation-defense-wenqian-liu/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210903T110000
DTEND;TZID=America/New_York:20210903T120000
DTSTAMP:20260426T130255
CREATED:20210825T175854Z
LAST-MODIFIED:20210825T175854Z
UID:5137-1630666800-1630670400@ece.northeastern.edu
SUMMARY:ECE PhD Proposal Review: Nikita Mirchandani
DESCRIPTION:PhD Proposal Review: Ultra-Low Power and Robust Analog Computing Circuits and System Design Framework for Machine Learning Applications \nNikita Mirchandani \nLocation: Zoom Link \nAbstract: As the scaling of CMOS transistors has almost halted\, performance gains of digital systems have also started to stagnate. There is a renewed interest in alternate computing techniques such as in-memory computing\, hybrid computing\, approximate computing\, and analog computing. In particular\, analog computing has reemerged as a promising alternative to save power and improve performance specifically for machine-learning (ML) applications. Analog computing has better area and power efficiency when compared to their digital counterpart. Power and chip area efficiency make analog computing highly appealing for implementing deep learning algorithms on-chip\, computing circuits for the internet-of-things (IoT) devices\, and implantable and wearable biomedical devices. However\, compared to digital computing\, analog computing methods have not nearly been utilized to their fullest potential due to longstanding challenges related to reliability\, programmability\, power consumption\, and high susceptibility to variations.\nThe subject of this dissertation research is to develop robust ultra-low power analog hardware suitable for machine learning applications. First\, a robust analog design methodology is presented to address issues of variability in analog circuits. A constant transconductance design technique using switched capacitor circuits is presented. The design approach is then applied to build circuits for ML applications. An analog vector matrix multiplier (VMM) is designed to be used in the convolutional layer in an ML analog computing vision hardware platform. Computing circuits are tested as part of an image classification DNN algorithm on the MNIST dataset and can achieve a classification accuracy of 96.1%. \nThe design approach is also used to design an analog computing system architecture for detection of seizures using EEG signals. A conventional EEG monitoring system includes an analog front-end (AFE)\, ADC\, digital filtering stage\, EEG feature extraction engine\, and SVM classification. Such systems suffer from high power and chip area requirements. The corresponding analog architecture is composed of AFE amplifiers to provide gain for the incoming signal. The AFE is followed by an analog filtering stage\, where spectral power from each of the bands is used as a feature for seizure classification. The output of each filter is applied to a corresponding feature extraction circuit to continuously monitor the onset of a seizure in an ultra-lower power mode with sub-threshold analog processing. The system level architecture is first modeled to obtain classification accuracy of seizures. Simulation times for the design of such complex analog systems can be prohibitively long\, particularly when the impacts of nonidealities such as noise\, nonlinearity\, and device mismatches have to be considered at the system level. The simulation time is reduced by building accurate models of the analog blocks for faster simulations. The analog models help to define the required specifications for each block in order to achieve a specified system-level classification accuracy.\nInfrastructure circuits like oscillators and voltage regulators for the proposed SoC are presented. A 254 nW 21 kHz on-chip RC oscillator with 21.5 ppm/oC temperature stability is presented to provide stable clock source for the proposed SoC.
URL:https://ece.northeastern.edu/event/ece-phd-proposal-review-nikita-mirchandani/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210830T093000
DTEND;TZID=America/New_York:20210830T103000
DTSTAMP:20260426T130255
CREATED:20210819T215252Z
LAST-MODIFIED:20210819T215252Z
UID:5129-1630315800-1630319400@ece.northeastern.edu
SUMMARY:ECE PhD Proposal Review: Keng Chen
DESCRIPTION:PhD Proposal Review: Design of Accurate and Responsive Power Management Integrated Circuits for Microprocessor Systems \nKeng Chen \nLocation: Zoom Link \nAbstract: As technology improves\, the processors used in data centers are becoming increasingly powerful. Since the number of complementary metal-oxide semiconductor (CMOS) transistors integrated into these processors continues to rise\, it becomes more challenging to design the power management integrated circuits to accurately regulate the supply voltages and ensure fast operational speeds for the processors. Furthermore\, with multi-core technology being extensively used in processor design\, there is more than just one voltage rail per chip that needs to be precisely regulated. In this research\, circuit design methods will be developed to improve the regulation accuracy\, both of the output voltage and converter switching frequency\, which are proposed for an integrated buck regulator designed for point of load (POL) applications. In order to ensure both accuracy and fast speed of operation\, a constant-on-time (COT) architecture was selected for this integrated buck regulator design approach. To provide accurate output voltage regulation\, the on-chip feedback sensing network and the bandgap reference circuit need to exhibit high accuracy. For this reason\, a highly accurate bandgap reference generation circuit with 5.8 ppm/°C – 13.5 ppm/°C over a wide operational region (-40 °C to 150 °C) has been designed. This bandgap reference circuit has a current-mode architecture and provides a 1.16V reference voltage with a 3.3V supply. A multi-section curvature compensation method is proposed to alleviate the nonlinear temperature-dependent error from the bipolar junction transistor’s base-emitter voltage. The two operational amplifiers utilized in this bandgap reference design to generate proportional-to-absolute-temperature (PTAT) and complementary-to-absolute-temperature (CTAT) current sources share one auxiliary auto-zero amplifier to ensure low input-referred offset voltage. Within many voltage regulators\, the feedback sensing network consists of multiple operational amplifiers\, and a major error source is the impact of the input-referred offset voltages of these amplifiers. \nThis research introduces the utilization of two different input-referred offset voltage correction methods for multiple amplifiers. The multi-amplifier system under investigation is used for feedback sensing during voltage regulation. It contains an instrumentation amplifier consisting of three folded cascode stages\, and an additional amplifier configured as a unity-gain buffer for a reference voltage. The first method in this work alleviates voltage offsets in a 4-amplifier system based on a shared auxiliary amplifier correction circuit that switches between different target amplifiers. The second method applies a chopping-based auto-zero procedure to cancel the input-referred offset voltage of the same system. Since chopping causes voltage ripples\, a correction circuit with a successive approximation register (SAR) analog-to-digital converter is used to reduce the output ripples. Both proposed methods can achieve less than 1mV feedback sensing error in voltage regulator applications.\nFor the constant-on-time buck regulator itself\, a new frequency locking loop to regulate the on-time pulse is also proposed in this work. The system shows less than 1% frequency shift over a wide programmable operational frequency range (400 KHz to 2 MHz). This frequency accuracy will further benefit the constant-on-time circuit to meet electromagnetic interference requirements without harming the fast-transient performance.
URL:https://ece.northeastern.edu/event/ece-phd-proposal-review-keng-chen/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210825T103000
DTEND;TZID=America/New_York:20210825T113000
DTSTAMP:20260426T130255
CREATED:20210818T000348Z
LAST-MODIFIED:20210823T174620Z
UID:5115-1629887400-1629891000@ece.northeastern.edu
SUMMARY:ECE PhD Dissertation Defense: Apoorve Mohan
DESCRIPTION:PhD Dissertation Defense: Rethinking the Choice between Static and Dynamic Provisioning for Centralized Bare-Metal Deployments \nApoorve Mohan \nLocation: Zoom Link \nAbstract: Technological inertia leads many organizations to continue to employ static provisioning strategies for bare-metal deployments. This results in architectural and runtime inefficiencies. However\, most performance and security-sensitive organizations currently employ static provisioning strategies to run and manage their workloads on physical servers (also known as bare-metal servers). This thesis demonstrates how recent technological advances enable dynamic provisioning strategies to improve aggregate bare-metal resource efficiency in centralized bare-metal deployments. Conceptually\, this dissertation has four parts. First\, it introduces a new system architecture for dynamic bare-metal provisioning. The proposed architecture reduces overhead and operational complexity and improves fault tolerance and performance. To achieve these improvements\, it decouples the provisioned state from the bare-metal servers. Second\, the thesis identifies the performance and operational issues due to intrusive software introspection strategies employed in existing bare-metal deployments. We present a dynamic bare-metal provisioning-based system to enable non-intrusive software introspection of bare-metal deployments to mitigate these issues. Many organizations have realized the cost benefits of consolidating their computing infrastructure over managing multiple\, relatively smaller compute facilities. \nOver the past decade\, virtual infrastructure solutions have become the obvious choice in consolidated (i.e.\, single-tenant managed or centralized) deployments for many organizations due to the scalability\, availability\, and cost benefits it offers. However\, most performance and security-sensitive organizations such as medical companies and hospitals\, financial institutions\, federal agencies run their workloads directly on physical (i.e.\, bare-metal) infrastructure. They are unwilling to tolerate the performance unpredictability and security risks due to co-location and performance overhead or vulnerabilities due to a large code base of the complex virtualization stack. In addition\, even if an organization uses virtual infrastructure\, bare-metal infrastructure is used to set up virtualization software stack and when their workloads require direct and exclusive access to hardware components (e.g.\, InfiniBand\, RAID\, FPGAs\, GPUs) that are difficult to virtualize. Such organizations invest enormous sums of money to buy or rent bare-metal servers and set up and manage bare-metal clusters. Thus\, it is imperative that these organizations efficiently operate and utilize the bare-metal clusters they set up to maximize their investment returns. However\, despite the technological advances over the past decade\, organizations continually employ static provisioning strategies in bare-metal deployments that contribute to poor aggregate bare-metal resource efficiency. Thesis Statement: This thesis demonstrates how dynamic provisioning can mitigate observed inefficiencies in centralized bare-metal deployments. The proposed dynamic provisioning strategies leverages existing storage disaggregation and fault-tolerance technologies to improve aggregate bare-metal resource efficiency. By applying it to four real-world scenarios\, this thesis demonstrates the effectiveness of dynamic provisioning.
URL:https://ece.northeastern.edu/event/ece-phd-dissertation-defense-apoorve-mohan/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210813T130000
DTEND;TZID=America/New_York:20210813T140000
DTSTAMP:20260426T130255
CREATED:20210812T222754Z
LAST-MODIFIED:20210812T222754Z
UID:5112-1628859600-1628863200@ece.northeastern.edu
SUMMARY:ECE MS Thesis Defense: Hafiyya Malik
DESCRIPTION:MS Thesis Defense: Studies for Improving Multi-channel Confocal Fluorescence Imagining in a Multi-modal Microscope \nHafiyya Malik \nLocation: Zoom Link \nMeeting ID: 988 8423 1362 Passcode: 399185 \nAbstract: The goal of this engineering thesis was to implement confocal fluorescence microscopy in an upgraded legacy electro-optical microscopy system for characterization of the viability of cells in biofilms. This application necessitated a 2-channel fluorescence measurement with the possible addition of confocal reflectance for context. Two major areas were addressed. First was the potential for using a silicon photomultiplier (SiPM) detector to replace the usual vacuum photomultiplier employed in confocal fluorescence microscope. A SiPM development board was purchased and tested to determine its performance at different light levels and in the presence of background light. The SiPM has the potential to be a more robust and reliable detector\, with the one disadvantage of a smaller area than some vacuum photomultipliers. The second area was the alignment of a long optical path that must include optics for complementary modes of imaging and allow for automated measurement across two spectral channels without loss of alignment. Alignment procedures were developed to simplify the process of alignment and a sliding stage was used to maintain alignment of two dichroic filters when changing channels. Based on this work\, multi-channel confocal fluorescence imaging in conjunction with confocal reflectance and bright-field microscope will be possible.
URL:https://ece.northeastern.edu/event/ece-ms-thesis-defense-hafiyya-malik/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210812T113000
DTEND;TZID=America/New_York:20210812T123000
DTSTAMP:20260426T130255
CREATED:20210809T173706Z
LAST-MODIFIED:20210810T185007Z
UID:5101-1628767800-1628771400@ece.northeastern.edu
SUMMARY:ECE Teaching Presentation: Mohammad Fanaei
DESCRIPTION:Title: Introduction to Python Programming with Applications in Computational Problem Solving. \nMohammad Fanaei \nLocation: Zoom Link \nAbstract: Python is a high-level\, interpreted programming language\, which supports multiple programming paradigms\, including functional and object-oriented programming. In this introductory learning conversation\, we will introduce the conditional execution of code\, loops\, and function definitions in Python. We will then apply the newly learned concepts to implement several numerical analysis methods and a Monte-Carlo simulation. The lecture will involve peer learning and assessment and assumes no background knowledge of Python programming. \nBio: Dr. Mohammad Fanaei received his Ph.D. degree in Electrical Engineering from West Virginia University\, Morgantown\, WV\, in 2016\, and his M.Sc. and B.Sc. degrees\, both in Electrical Engineering\, from Isfahan University of Technology\, Iran\, in 2008 and 2005\, respectively. Over the last six years\, Dr. Fanaei has worked at three different universities\, as an assistant professor of Electrical and Robotics & Mechatronic Systems Engineering at the University of Detroit Mercy (from 2016 to 2017 and from 2018 to present)\, as an assistant professor of Electrical Engineering at Bucknell University (from 2017 to 2018) and in the Iron Range Engineering Program in the Department of Integrated Engineering at Minnesota State University Mankato (from 2015 to 2016). Dr. Fanaei’s research interests are in the broad areas of the design\, analysis\, and evaluation of machine learning and deep learning technologies enabling connected\, automated\, and autonomous driving systems\, as well as the applications of stochastic signal processing in wireless communication systems and sensor networks. His teaching interests include embedded systems\, digital design\, wireless networks\, cryptology and network security\, communication systems\, stochastic signal processing\, and digital signal processing.
URL:https://ece.northeastern.edu/event/ece-teaching-presentation-mohammad-fanaei/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210810T100000
DTEND;TZID=America/New_York:20210810T110000
DTSTAMP:20260426T130255
CREATED:20210727T192945Z
LAST-MODIFIED:20210727T192945Z
UID:5077-1628589600-1628593200@ece.northeastern.edu
SUMMARY:ECE PhD Dissertation Defense: Pu Zhao
DESCRIPTION:PhD Dissertation Defense: Towards Robust Image Classification with Deep learning and Real-Time DNN Inference on Mobile \nPu Zhao \nLocation: Zoom Link \nAbstract:  As the rapidly increasing popularity of deep learning\, deep neural networks (DNN) have become the fundamental and essential building blocks in various applications such as image classification and object detection. However\, there are two main issues which potentially limit the wide application of DNNs: 1) the robustness of DNN models raises security concerns\, and 2) the large computation and storage requirements of DNN models lead to difficulties for its wide deployment on popular yet resource-constrained devices such as mobile phones.\nTo investigate the DNN robustness\, we explore the DNN attack\, robustness evaluation and defense. More specifically\, for DNN attack\, we achieve various attack goals (e.g. adversarial examples and fault sneaking attacks) with different algorithms (e.g. alternating direction method of multipliers (ADMM) and natural gradient descent (NGD) attacks) under various conditions (white-box and black-box attacks). For robustness evaluation\, we propose a fast evaluation method to obtain the model perturbation bound such that any model perturbation within the bound does not alter the model classification outputs or incur model mis-behaviors. For the DNN defense\, we investigate the defense performance with model connection techniques and successfully mitigate the fault sneaking and backdoor attacks. With a deeper understanding of the DNN robustness\, we further explore the deployment problem of DNN models on edge devices with limited resources.\nTo satisfy the storage and computation limitation on edge devices\, we adopt model pruning to remove the redundancy in models\, thus reducing the storage and computation during inference. Besides\, as some applications have real-time requirements with high inference speed sensitivities such as object detection on autonomous cars\, we further try to implement real-time DNN inference for various DNN applications on mobile devices with pruning and compiler optimization. To summary\, we mainly investigate the DNN robustness and implement real-time DNN inference on the mobile.
URL:https://ece.northeastern.edu/event/ece-phd-dissertation-defense-pu-zhao/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210806T140000
DTEND;TZID=America/New_York:20210806T150000
DTSTAMP:20260426T130255
CREATED:20210730T184341Z
LAST-MODIFIED:20210730T184341Z
UID:5091-1628258400-1628262000@ece.northeastern.edu
SUMMARY:ECE PhD Dissertation Defense: Mohammadreza Sharif
DESCRIPTION:PhD Dissertation Defense: Human-in-the-loop Prosthetic Robot Hand Control through Contextual Adaptation \nMohammadreza Sharif \nLocation: ISEC136 or Zoom Link \nAbstract: Despite fifty years of research on prosthetic robot hands\, this technology is yet to be fully acknowledged by amputees as well as manufacturers\, due to lack of robustness and intuitive control. Our goal is to enhance the control robustness and intuitiveness through integrating the context\, i.e. knowledge of environment and task\, with human bio-signals\, e.g. hand trajectory. Although this solution has already been studied in the literature\, no unified framework is proposed for multi-modal information fusion. This research is aimed at introducing a novel framework for human-in-the-loop prosthetic robot hand control. We propose our solution in two parts\, (1) grasp inference and (2) end-to-end actuator control. For (1)\, we propose a model-based and a model-free grasp inference framework. Our model-based method is based on particle filters method. With knowledge of context hard-coded into the solution per se\, our particle-filter-based framework can incorporate any input signal using a proper weight function. In our model-free method we use hidden Markov mode to learn the grasp-inference task directly from human hand transport trajectories. For (2)\, we propose a reinforcement learning (RL) framework which learns to control robot actuator directly from the context with less information hard-coded into the solution. We leverage imitation learning (IL) besides RL to overcome challenging exploration in the problem. To provide invariance to the human hand transport trajectories\, we first provide a solution based on synthesized trajectories based on a human motion model. Later\, we adopt an over-sampling technique for real human hand transport trajectories\, to serve as a means of data augmentation. This research provides a step forward in more rigorous frameworks for multi-modal information fusion for prosthetic robot hand control and grasp inference through model- /data-based methods. \n 
URL:https://ece.northeastern.edu/event/ece-phd-dissertation-defense-mohammadreza-sharif/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210804T130000
DTEND;TZID=America/New_York:20210804T140000
DTSTAMP:20260426T130255
CREATED:20210727T225554Z
LAST-MODIFIED:20210727T225554Z
UID:5081-1628082000-1628085600@ece.northeastern.edu
SUMMARY:ECE MS Thesis Defense: Hooman Barati Sedeh
DESCRIPTION:MS Thesis Defense: Space-Time Graphene Metasurfaces \nHooman Barati Sedeh \nLocation: Zoom Link \nAbstract: The unprecedented growth of the data exchanged between wireless devices and the rapid emergence of high-quality wireless services have raised the demand for communication bandwidth and data transmission rates. This has motivated the migration of wireless networks toward the utilization of carrier waves with higher frequencies beyond the millimeter-wave band. Terahertz (THz) band is envisioned as one of the enabling technologies for future generations of wireless communication mobile networks (such as 6G) to address the needs of high-speed and bandwidth-intensive applications. THz communications are expected to provide broadband services to several wireless devices in the internet of things applications. The recent migration from internet-of-things to the internet of nano-things imposes certain constraints in terms of size\, weight\, and power (SWaP) on the THz antennas responsible for communications\, calling for the development of smart antennas with adaptive response capable of establishing multiple active data links through multi-beam scanning in a multiple-input\, multiple-output (MIMO) network while meeting the demands on the capacity and SWaP. Moreover\, provisioning a reliable communication channel that ensures the security of its users’ information remains vitally essential for the next generation of communication networks.\nMetasurfaces\, consisting of subwavelength elements\, are poised to enable improved free-space optical communications with low SWaP thanks to their small form factor and capability to provide unprecedented control over the wavefront of electromagnetic waves at the subwavelength scale. The recent investigations of active metasurfaces have aimed toward overcoming the fixed response of conventional metasurfaces and developing smart antenna systems with adaptive beamforming capabilities that can point the beam toward the desired users in real-time through pixelated control over the phase of the scattered wave. Moreover\, the adaptive communication by such quasi-static tunable metasurfaces can be secured by encrypting the transmitted data via holography to impose restrictions on the data access from an adversary. Despite the fruitful progress in this area\, quasi-static active metasurface face several challenges to meet the high demands on the capacity of communication due to their reliance on resonant phase shift accumulations which limit the operation bandwidth and hinders the scalability in terms of the number of channels in the account of non-trivial coupling effects between resonant unit cells. Furthermore\, these metasurfaces cannot be used for covert communication as they do not allow for engineering the spectral content of scattered light.\nThis thesis explores the roles of space and time in active metasurfaces for establishing adaptive and secure multichannel communication at low-THz frequency regime. As the primary goal of this work is twofold\, we will tackle each problem separately. At first\, we propose a technique for adaptive multichannel communication through simultaneous and independent multifrequency multibeam scanning via a single time-modulated metasurface consisting of graphene micro-patch antennas whose Fermi energy levels are modulated by radio-frequency biasing signals. To this aim\, we divide the metasurface aperture into interleaved orthogonally modulated sub-array antennas with distinct modulation frequencies\, rendering a shared aperture in space-time. The higher-order frequency harmonics generated by the sub-arrays in such a space-time shared-aperture metasurface are mutually orthogonal in the sense that they do not yield an observable interference pattern and can be separated by spectral filtering. A distinct constant progressive modulation phase delay is then adopted in each sub-array to independently scan its corresponding higher-order frequency harmonics via dispersionless modulation-induced phase gradient with minimal sidelobe level and full angle-of-view over a wide bandwidth. In the second part of this work\, we will propose another technique for establishing active secure communication links over single and multiple orthogonal frequency channels via a metasurface that consists of graphene micro-ribbons. To this aim\, the Fermi energy level of each graphene micro-ribbons is modulated via pseudo-random radio-frequency biasing signals whose DC offsets are adjusted to tilt the reflected beam toward the predefined direction by imposing a spatial phase gradient profile across the surface\, while their waveforms are engineered to expand the incident wave spectrum into a noise-like spectrum with a near-zero power spectral density via random modulation of the reflection phase of each element with respect to its offset phase. This permits for addressing a legitimate mobile user in real-time who can retrieve the incident signal via synchronous demodulation with the pseudo-random key of the metasurface while camouflaging the signal from the adversary by lowering the probability of detection and spectral encryption. The approach is then extended to enable multi-channel secure communication by dividing the metasurface into interleaved sub-arrays modulated with orthogonal pseudo-random keys\, which provides simultaneous and independent control over multiple beams with non-overlapping spread spectra which can be retrieved by independent legitimate users while rejecting unwarranted access by eavesdroppers as well as other users.
URL:https://ece.northeastern.edu/event/ece-ms-thesis-defense-hooman-barati-sedeh/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210804T100000
DTEND;TZID=America/New_York:20210804T110000
DTSTAMP:20260426T130255
CREATED:20210729T182123Z
LAST-MODIFIED:20210729T182123Z
UID:5085-1628071200-1628074800@ece.northeastern.edu
SUMMARY:ECE MS Thesis Defense: Owen McElhinney
DESCRIPTION:MS Thesis Defense: On the Application of Spline Functions to Problems in Hyperspectral Imaging \nOwen McElhinney \nLocation: Zoom Link \nAbstract: Hyperspectral Imaging (HSI) is a rapidly growing topic in the field of remote sensing. Hyperspectral cameras trade off a reduction in the spatial resolution of modern imaging for a higher spectral resolution. This allows for the detection of surface materials using the principles of spectroscopy.\nThis work will investigate the application of a class of functions called splines to three different problems in the field of HSI. Splines are a special class of data-fitting functions that guarantee continuity. Common data-fitting techniques like polynomial and piecewise-polynomial fitting are unable to match the complexity of HSI data. Splines provide a robust fitting procedure that matches the physical reality of material spectra. They can additionally be used to smooth data\, calculate derivatives\, interpolate between points\, and more.\nThe first problem will look at the smoothing of noisy data for detecting small materials. When objects are smaller than the pixel\, the observed spectrum will mix the target with background materials. The problem of unmixing removes the influence of these background materials from observed data. If the object is too small\, the estimates from unmixing will be dominated by error. In this scenario\, splines will be used to smooth out these random variations.\nThe second problem will use splines to introduce a new solution to the problem of Temperature Emissivity Separation (TES). The physical quantity captured by the camera is radiance. For the detection of materials\, ground reflectance or emissivity are desired. TES is the process by which ground radiance is converted to material emissivity. Splines will be used to replace estimated roughness in this problem with an analytical solution.\nThe third and final application looks at using splines to detect gases without relying on image statistics. Gas features are sharp and only impact a narrow window of the spectrum. This application attempted to use splines to detect these sharp features by looking at the difference between the collected data and an interpolation across the feature.\nThe first two applications yielded interesting and useful results. The third application yielded some interesting conclusions about the general problem and improved methods for using splines in this space.
URL:https://ece.northeastern.edu/event/ece-ms-thesis-defense-owen-mcelhinney/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210804T100000
DTEND;TZID=America/New_York:20210804T110000
DTSTAMP:20260426T130255
CREATED:20210727T225426Z
LAST-MODIFIED:20210727T225426Z
UID:5079-1628071200-1628074800@ece.northeastern.edu
SUMMARY:ECE PhD Dissertation Defense: Amirreza Farnoosh
DESCRIPTION:PhD Dissertation Defense: Unsupervised Learning of Low-Dimensional Dynamical Representations from Spatiotemporal Data \nAmirreza Farnoosh \nLocation: Zoom Link \nAbstract: Ever-improving sensing technologies offer a fast and accurate collection of large-scale spatiotemporal data\, recorded from multimodal sensors of heterogeneous natures\, in various application domains\, ranging from medicine and biology to robotics and traffic control. In this dissertation\, we propose frameworks for learning the underlying representation of these data in an unsupervised manner\, tailored towards several emerging applications\, namely indoor navigation and mapping\, neuroscience hypothesis testing\, time series forecasting\, 3D motion segmentation\, and human action recognition.\nAs such\, (1) we developed an unsupervised framework for real-time depth and view-angle estimation from an inertially augmented video recorded from an indoor scene by employing geometric-based machine learning and deep learning models. (2) We introduced a hierarchical deep generative factor analysis framework for temporal modeling of neuroimaging datasets. Our model approximates high dimensional data by a product between time-dependent weights and spatially-dependent factors which are in turn represented in terms of lower dimensional latents. This framework can be extended to perform clustering in the low dimensional temporal latent or perform factor analysis in the presence of a control signal. (3) We developed a deep switching dynamical system for dynamical modeling of multidimensional time-series data. Specifically\, we employ a deep vector auto-regressive latent model switched by a chain of discrete latents to capture higher-order multimodal latent dependencies. This results in a flexible model that (i) provides a collection of potentially interpretable states abstracted from the process dynamics\, and (ii) performs short- and long-term vector time series prediction in a complex multi-relational setting. (4) We developed a dynamical deep generative latent model for segmentation of 3D pose data over time that parses the meaningful intrinsic states in the dynamics of these data and enables a low-level dynamical generation and segmentation of skeletal movements. Our model encodes highly correlated skeletal data into a set of few spatial basis of switching temporal processes in a low-dimensional latent framework. We extended this model for human action recognition by decoding from these low-dimensional latents to the motion data and their associated action labels.
URL:https://ece.northeastern.edu/event/ece-phd-dissertation-defense-amirreza-farnoosh/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210802T140000
DTEND;TZID=America/New_York:20210802T150000
DTSTAMP:20260426T130255
CREATED:20210729T185822Z
LAST-MODIFIED:20210729T185822Z
UID:5087-1627912800-1627916400@ece.northeastern.edu
SUMMARY:ECE PhD Dissertation Defense: Yulun Zhang
DESCRIPTION:PhD Dissertation Defense: Deep Convolutional Neural Network for Image Restoration and Synthesis \nYulun Zhang \nLocation: Zoom Link \nAbstract: Image restoration and synthesis with deep learning play a fundamental role in the computer vision community. They are widely used on mobile devices (e.g.\, smartphones) or lead to billion-dollar startups. However\, how to design efficient deep convolutional neural networks (CNNs) to extract higher-quality deep CNN features for better image restoration and synthesis is still challenging. In this dissertation talk\, I will describe my recent works to enhance CNN features in the channel dimension or/and the spatial dimensions. First\, for image restoration\, I will briefly introduce our proposed residual dense network. Then\, I will introduce the residual in residual (RIR) structure to train very deep super-resolution networks. Such an RIR structure could also make the network learn more high-frequency information\, being critical for high-resolution output. Attention mechanism (e.g.\, channel attention and spatial attention) is further explored to highlight the features. Second\, for image synthesis\, I will introduce multimodal style transfer via graph cuts. I visualize the deep features and find the multimodal style representation. I then formulate the style matching problem as an energy minimization one\, which could be solved via graph cuts. As a result\, the transferred features contain spatially semantic information\, providing more visually pleasing stylized results. Besides\, we investigate image synthesis about texture hallucination with large scaling factors. We propose an efficient high-resolution hallucination network for very large scaling factors.
URL:https://ece.northeastern.edu/event/ece-phd-dissertation-defense-yulun-zhang/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210729T150000
DTEND;TZID=America/New_York:20210729T160000
DTSTAMP:20260426T130255
CREATED:20210727T191933Z
LAST-MODIFIED:20210727T191933Z
UID:5075-1627570800-1627574400@ece.northeastern.edu
SUMMARY:ECE PhD Proposal Review: Weite Zhang
DESCRIPTION:PhD Proposal Review: High Sensing-capacity Multi-dimensional-coded Millimeter-wave MIMO Imaging System \nWeite Zhang \nLocation: Microsoft Teams Link \nAbstract: Millimeter-wave (mm-wave) MIMO imaging systems have been explored to use more and more complicated radar waveforms to achieve advanced multiplexing and high-performance imaging. As the complexity of the radar waveform increases\, conventional systems inevitably suffer from higher design difficulty and cost. In spite of the radar waveform design\, existing mm-wave imaging systems are still suboptimal due to the fact that the sensing matrix is not tailored properly to achieve its maximum capacity\, which often results in large mutual information between successive measurements\, and limited imaging performance.\nAs the first contribution of this proposal\, high sensing-capacity mm-wave MIMO imaging systems with multi-dimensional-coding are built. In the first prototype\, a 70-77 GHz frequency-modulated continuous wave (FMCW) MIMO imaging system with massive channels is studied. To enhance the sensing-capacity\, a compressive reflector antenna (CRA) is added to perform randomized spatial wavefront coding to increase the measurement diversity. Both static and on-the-move experiments are carried out to show the functionality of the imaging system. In the second prototype\, an 81-86 GHz software-defined mm-wave MIMO imaging system is designed\, which makes use of cost-effective software-defined radios (SDRs) with mm-wave mixers. Due to the baseband flexibility of SDRs\, efficient orthogonal frequency-division multiplexing (OFDM) with binary phase coding is designed as the radar waveform to achieve simultaneous MIMO transmission\, where high receiving signal-to-noise ratio and spectrum efficiency are achieved. Again\, a CRA is designed and applied to increase the measurement diversity. Primary simulation and experimental results show good imaging performance with reduced side lobe effect.\nAs the second contribution of this proposal\, a material characterization method is developed\, which is vital in some important mm-wave imaging applications\, such as security screening\, where both object profile and material information are required for potential threats prediction. Specifically\, a Geometrical Optics (GO) forward model based on a reflectarray imaging system is developed. The GO forward model can be adapted to any other imaging systems as long as their geometrical configurations are known. Both simulations and experiments are performed to show the effectiveness and efficiency of the proposed material characterization method\, where the complex relative permittivity as well as a more accurate shape of the object is retrieved.
URL:https://ece.northeastern.edu/event/ece-phd-proposal-review-weite-zhang/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210728T140000
DTEND;TZID=America/New_York:20210728T150000
DTSTAMP:20260426T130255
CREATED:20210727T191651Z
LAST-MODIFIED:20210727T191651Z
UID:5073-1627480800-1627484400@ece.northeastern.edu
SUMMARY:ECE PhD Proposal Review: Bahare Azari
DESCRIPTION:PhD Proposal Review: Circular-Symmetric Correlation Layer based on FFT \nBahare Azari \nLocation: Zoom Link \nAbstract: Planar convolutional neural networks\, widely known as CNNs\, have been exceptionally successful in many computer vision and machine learning tasks\, such as object detection\, tracking\, and classification. The convolutional layers in CNN are characterized by pattern-matching filters that can identify motifs in the signal residing on a 2D plane. However\, there exists various applications in which we have signals lying on a curved manifold or an arbitrary collection of coordinates\, e.g.\, temperature and climate data on the surface of the (spherical) earth\, and 360-panoramic images acquired from LiDAR. In these applications\, we usually need our network to be equivariant/invariant to various transformations of the input\, i.e.\, as we transform the input according to a certain action of a group\, the output is respectively transformed (equivariance)\, or remains unchanged (invariance). The convolution layers are empirically known to be invariant to small translations of their input image\, but they are not completely immune to relatively large translations Hence\, they may fail on the tasks that requires invariance to a specific transformation\, and and on the data that includes a wide range of that transformation. \nIn this work we consider equivariant/invariant tasks on 360-panoramic data. For a systematic treatment of analyzing the 360-panoramic data\, we propose a circular-symmetric correlation Layer (CCL) based on the formalism of roto-translation equivariant correlation on the continuous group constructed of the unit circle and the real line. We implement this layer efficiently using the well-known Fast Fourier Transform (FFT) and discrete cosine transform (DCT) algorithm. We discuss how the FFT yields the exact calculation of the correlation along the panoramic direction due to the circular symmetry and guarantees the invariance with respect to circular shift. The DCT provides an improved approximation with respect to transnational symmetry compared to what we observe in CNNs. We demonstrated the invariance analysis of networks built with CCL on two benchmark datasets comparing the equivariance of neural networks adopting CCL layers and regular CNN. Then\, we showcase the performance analysis of a general network equipped with CCL on recognition and classification tasks\, such as panoramic scene change detection\, 3D object classification\, LIDAR Semantic Segmentation.
URL:https://ece.northeastern.edu/event/ece-phd-proposal-review-bahare-azari/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210723T100000
DTEND;TZID=America/New_York:20210723T110000
DTSTAMP:20260426T130255
CREATED:20210707T005406Z
LAST-MODIFIED:20210707T005406Z
UID:5032-1627034400-1627038000@ece.northeastern.edu
SUMMARY:ECE PhD Dissertation Defense: Mahmoud Ibrahim
DESCRIPTION:PhD Dissertation Defense: Low-Power Integrated Circuit Design for Wireless Devices in the Internet of Things \nMahmoud Ibrahim \nLocation:  Zoom \nAbstract: Numerous integrated sensing devices are under development for wireless medical diagnostic and monitoring applications. However\, the data rates of wireless devices connected to the Internet of Things are limited and strongly depend on the available power. This research addresses the need for circuit-level design methods to enable higher data rates with lower power consumption in order to facilitate the proliferation of wireless devices that can overcome the speed-power conundrum. The potential applications include continuous-time monitoring of physiological signals\, where increased data rates imply the ability to exchange more information during the same time\, more accurate data\, and/or data from a greater number of sites associated with each wireless node.\nAn energy-efficient binary frequency shift keying (BFSK) transmitter architecture for biomedical applications is introduced as the first part of this dissertation research. To achieve low power consumption with higher data rates\, the novel transmitter architecture leverages image rejection techniques to generate each of the two tones of the transmitted BFSK signal while keeping the phase-locked loop (PLL) oscillator frequency unchanged\, and thus maintaining low PLL power and overall transmitter power. A fabricated prototype chip in 130nm complementary metal-oxide-semiconductor (CMOS) technology achieves data rates up to 10 Mbps while consuming 180 µW with up to -20 dBm output power according to Medical Implant Communication System (MICS) band requirements. The measurement results confirm state-of-the-art energy-efficient performance with 18 pJ/bit.\nAs a natural continuation of the first part of this research\, a complementary receiver architecture is described in the second part of this dissertation to provide full transceiver capabilities. The new receiver design approach takes advantage of the transmitted signal characteristics by using both the frequency information and phase information to demodulate the received digital bits. This design method results in improved sensitivity with reduced power consumption through relaxed receiver block specification requirements. The custom-designed receiver circuits include a new low-noise amplifier (LNA) topology for energy-efficient antenna impedance matching\, and a single mixer circuit that realizes the signal down-conversion with differential in-phase and quadrature-phase baseband output signals to circumvent the complexity associated with two mixers and to save power. Measurement results of the fabricated receiver in 65nm CMOS technology show a sensitivity of -82 dBm with an input signal at 10 Mbps centered around 416 MHz. With a power consumption of 610 µW and an energy efficiency of 61 pJ/bit\, this receiver architecture displays state-of-the-art performance with respect to data rate\, power and sensitivity compared to other receivers in the same frequency range.\nIn addition to the new transmitter and receiver architectures\, a large-signal transconductance linearization technique is presented as part of this dissertation research to extend the dynamic range of analog baseband filters. Furthermore\, a low-power sinusoidal signal generation technique is introduced and analyzed\, which is a versatile and essential component of the transmitter design approach.
URL:https://ece.northeastern.edu/event/ece-phd-dissertation-defense-mahmoud-ibrahim/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210719T113000
DTEND;TZID=America/New_York:20210719T123000
DTSTAMP:20260426T130255
CREATED:20210713T212719Z
LAST-MODIFIED:20210713T212719Z
UID:5046-1626694200-1626697800@ece.northeastern.edu
SUMMARY:ECE PhD Dissertation Defense: Berkan Kadioglu
DESCRIPTION:PhD Dissertation Defense: An Analysis of Algorithms with Discrete Choice Models \nBerkan Kadioglu \nLocation: Zoom Link \nAbstract: In the first half of our work\, we consider a rank regression setting\, in which a dataset of $N$ samples with features in $\mathbb{R}^d$ is ranked by an oracle via $M$ pairwise comparisons.\nSpecifically\, there exists a latent total ordering of the samples; when presented with a pair of samples\, a noisy oracle identifies the one ranked higher w.r.t. the underlying total ordering.\nA learner observes a dataset of such comparisons\, and wishes to regress sample ranks from their features.\nWe show that to learn the model parameters with $\epsilon > 0$ accuracy\, it suffices to conduct $M \in \Omega(dN\log^3 N/\epsilon^2)$ comparisons uniformly at random when $N$ is $\Omega(d/\epsilon^2)$.\nCompared to learning from class labels\, learning from comparison labels has two advantages: First\, comparison labels reveal both inter and intra-class information\, where class labels only contain the former.\nSecond\, comparison labels also exhibit lower variability across different labelers.\nThis has been observed experimentally in multiple domains\, including medicine \citep{campbell2016plus\,kalpathy2016plus\, stewart2005absolute} and recommendation systems \citep{schultz2004learning\,zheng2009mining\,brun2010towards\, koren2011ordrec}\, and is due to the fact that humans often find it easier to make relative\, rather than absolute\, judgements.\nMany works focusing on empirically learning comparison labels show excellent performance in practice \citep{tian2019severity\,yildiz2019classification}.\nOur work provides a theoretical foundation for analyzing and understanding this empirical performance.\nMoreover\, we extend the problem we initially study to a harder setting.\nWe do this by moving from pairwise comparisons to multi-way comparisons.\nFurthermore\, we study an online variant of the previous problem where the goal is to maintain high user engagement throughout the learning period.\nThis of course\, indirectly leads to the goal of learning parameters of the discrete choice model as accurately as possible\, fast.\nThis new problem is directly related to a setting in which a retailer recommends products to customers.\nA common problem in many recommendation tasks is to simultaneously learn the utilities of items to be recommended and maintain high user engagement.\nWe are generally constrained by a limit on the total number of items to be recommended at a time for an unknown time horizon.\nRecently\, bandit algorithms have been proposed for this setting where the multinomial logit model is assumed.\nBounds on error metrics are provided for upper confidence and Thompson sampling based algorithms.\nIn our paper\, we propose a variational inference based Thompson sampling algorithm and identify the required properties to achieve $\tilde O(D^{3/2}\sqrt T)$ worst-case regret.\nThrough extensive experiments we show that our method performs much better than the recently proposed \emph{TSMNL} algorithm in many error metrics.\nWe further accelerate our algorithm to be used in practical settings.
URL:https://ece.northeastern.edu/event/ece-phd-dissertation-defense-berkan-kadioglu/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210713T100000
DTEND;TZID=America/New_York:20210713T110000
DTSTAMP:20260426T130255
CREATED:20210706T174832Z
LAST-MODIFIED:20210706T174832Z
UID:5023-1626170400-1626174000@ece.northeastern.edu
SUMMARY:ECE PhD Dissertation Defense: Maher Kachmar
DESCRIPTION:PhD Dissertation Defense: Active Resource Partitioning and Planning for Storage Systems using Time Series Forecasting and Machine Learning Techniques \nMaher Kachmar \nLocation: Zoom \nAbstract: In today’s enterprise storage systems\, supported data services such as snapshot delete or drive rebuild can result in tremendous performance overhead if executed inline along with heavy foreground IO\, often leading to missing Service Level Objectives (SLOs). Moreover\, new classes of data services\, such as thin provisioning\, instant volume snapshots\, and data reduction features make capacity planning and drive wear-out prediction quiet challenging. Having enough free storage pool capacity available ensures that the storage system operates in favorable conditions during heavy foreground IO cycles. This enables the storage system to defer background work to a future idle cycle. Static partitioning of storage systems resources such as CPU cores or memory caches may lead to missing data reduction rate (DRR) guarantees. However\, typical storage system applications such as Virtual Desktop Infrastructure (VDI) or web services follow a repetitive workload pattern that can be learned and/or forecasted. Learning these workload pattern allows us to address several storage system resource partitioning and planning challenges that may not be overcome with traditional manual tuning and primitive feedback mechanism.\nFirst\, we propose a priority-based background scheduler that learns this pattern and allows storage systems to maintain peak performance and meet service level objectives (SLOs) while supporting a number of data services. When foreground IO demand intensifies\, system resources are dedicated to service foreground IO requests. Any background processing that can be deferred is recorded to be processed in future idle cycles\, as long as our forecaster predicts that the storage pool has remaining capacity. A smart background scheduler can adopt a resource partitioning model that allows both foreground and background IO to execute together\, as long as foreground IOs are not impacted\, harnessing any free cycles to clear background debt. Using traces from VDI and web services applications\, we show how our technique can out-perform a static policy that sets fixed limits on the deferred background debt and reduces SLO violations from 54.6% (when using a fixed background debt watermark)\, to only 6.2% when employing our dynamic smart background scheduler.\nSecond\, we propose a smart capacity planning and recommendation tool that ensures the right number of drives are available in the storage pool in order to meet both capacity and performance constraints\, without over-provisioning storage. Equipped with forecasting models that characterize workload patterns\, we can predict future storage pool utilization and drive wear-outs. Similarly\, to meet SLOs\, the tool recommends expanding pool space in order to defer more background work through larger debt bins. Overall\, our capacity planning tool provides a day/hour countdown for the next Data Unavailability/Data Loss (DU/DL) event\, accurately predicting DU/DL events to cover a future 12-hour time window.\nMoreover\, supported services such as data deduplication are becoming a common feature adopted in the data center\, especially as new storage technologies mature. Static partitioning of storage system resources\, memory caches\, may lead to missing SLOs\, such as the Data Reduction Rate (DRR) or IO latency. Lastly\, we propose a Content-Aware Learning Cache (CALC) that uses online reinforcement learning models (Q-Learning\, SARSA and Actor-Critic) to actively partition the storage system cache between a deduplicated data digest cache\, content cache\, and address-based data cache to improve cache hit performance\, while maximizing data reduction rates. Using traces from popular storage applications\, we show how our machine learning approach is robust and can out-perform an iterative search method for various data-sets and cache sizes. Our content-aware learning cache improves hit rates by 7.1% when compared to iterative search methods\, and 18.2\% when compared to traditional LRU-based data cache implementation.
URL:https://ece.northeastern.edu/event/ece-phd-dissertation-defense-maher-kachmar/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210707T170000
DTEND;TZID=America/New_York:20210707T180000
DTSTAMP:20260426T130255
CREATED:20210706T175131Z
LAST-MODIFIED:20210706T175131Z
UID:5027-1625677200-1625680800@ece.northeastern.edu
SUMMARY:ECE PhD Dissertation Defense: Kaidi Xu
DESCRIPTION:PhD Dissertation Defense: Can We Trust AI? Towards Practical Implementation and Theoretical Analysis in Trustworthy Machine Learning \nKaidi Xu \nLocation: Zoom Link \nAbstract: Deep learning has achieved extraordinary performance in many application domains recently. It has been well accepted that DNNs are vulnerable to adversarial attacks\, which raises concerns of DNNs in security-critical applications and may result in disastrous consequences. Adversarial attacks are usually implemented by generating adversarial examples\, i.e.\, adding sophisticated perturbations onto benign examples\, such that adversarial examples are classified by the DNN as target (wrong) labels instead of the correct labels of the benign examples. The adversarial machine learning aims to study this phenomenon and leverage it to build robust machine learning systems and explain DNNs.\nIn this talk\, I will present the mechanism of adversarial machine learning in both empirical and theoretical ways. Specifically\, a uniform adversarial attack generation framework\, structured attack (StrAttack) is introduced\, which explores group sparsity in adversarial perturbations by sliding a mask through images aiming for extracting key spatial structures. Second\, we discuss the feasibility of adversarial attacks in the physical world and introduce a convincing framework\, Expectation over Transformation (EoT). Utilize EoT with Thin Plate Spline (TPS) transformation\, we can generate Adversarial T-shirts\, a powerful physical adversarial patch for evading person detectors even if it could undergo non-rigid deformation due to a moving person’s pose changes. Third\, we stand on the defense side and design the first adversarial training method based on Graph Neural Network. Finally\, we introduce Linear relaxation-based perturbation analysis (LiRPA) for neural networks\, which computes provable linear bounds of output neurons given a certain amount of input perturbation. LiRPA studies the adversarial example in a theoretical way and can guarantee the test accuracy of a model by given perturbation constraints. The generality\, flexibility\, efficiency and ease-of-use of our proposed framework facilitate the adoption of LiRPA based provable methods for other machine learning problems beyond robustness verification
URL:https://ece.northeastern.edu/event/ece-phd-dissertation-defense-kaidi-xu/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210707T140000
DTEND;TZID=America/New_York:20210707T150000
DTSTAMP:20260426T130255
CREATED:20210706T175010Z
LAST-MODIFIED:20210706T175010Z
UID:5025-1625666400-1625670000@ece.northeastern.edu
SUMMARY:ECE PhD Proposal Review: Xiaolong Ma
DESCRIPTION:PhD Proposal Review: Towards Efficient Deep Neural Network Execution with Model Compression and Platform-specific Optimization \nXiaolong Ma \nLocation: Zoom \nAbstract: Deep learning or deep neural networks (DNNs) have become the fundamental element and core enabler of ubiquitous artificial intelligence. Recently\, with the emergence of a spectrum of high-end mobile devices\, many deep learning applications that formerly required desktop-level computation capability are being transferred to these devices. However\, executing DNN inference is still challenging considering the high computation and storage demands\, specifically\, if real-time performance with high accuracy is needed. Weight pruning of DNNs is proposed\, but existing schemes represent two extremes in the design space: non-structured pruning is fine-grained\, accurate\, but not hardware friendly; structured pruning is coarse-grained\, hardware-efficient\, but with higher accuracy loss. To solve the problem\, we propose a compression-compilation co-optimization framework\, which includes 1) a new dimension\, fine-grained pruning patterns inside the coarse-grained structures that achieves accuracy enhancement and preserve the structural regularity that can be leveraged for hardware acceleration\, 2) a pattern-aware pruning framework that achieves pattern library extraction\, pattern selection\, pattern and connectivity pruning and weight training simultaneously\, and 3) a set of thorough architecture-aware compiler/code generation-based optimizations\, i.e.\, filter kernel reordering\, compressed weight storage\, register load redundancy elimination\, and parameter auto-tuning for real-time execution of the mainstream DNN applications on the mobile platforms. Evaluation results demonstrate that our framework outperforms three state-of-the-art end-to-end DNN frameworks\, TensorFlow Lite\, TVM\, and Alibaba Mobile Neural Network with speedup up to 44.5x\, 11.4x\, and 7.1x\, respectively\, with no accuracy compromise. Real-time inference of representative large-scale DNNs (e.g.\, VGG-16\, ResNet-50) can be achieved using mobile devices.
URL:https://ece.northeastern.edu/event/ece-phd-proposal-review-xiaolong-ma/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210701T110000
DTEND;TZID=America/New_York:20210701T120000
DTSTAMP:20260426T130255
CREATED:20210623T211449Z
LAST-MODIFIED:20210623T211449Z
UID:5007-1625137200-1625140800@ece.northeastern.edu
SUMMARY:ECE PhD Dissertation Defense: Xianfeng Liang
DESCRIPTION:PhD Dissertation Defense: RF Magnetoelectric Microsystems \nXianfeng Liang \nLocation: Zoom Link \nAbstract: Multiferroic materials are the materials that inherently exhibit two or more ferroic properties\, such as ferroelectricity\, ferromagnetism and ferroelasticity\, etc. Magnetoelectric (ME) materials with coupled magnetization and electric polarization have attracted intense interests recently due to the realization of strong ME coupling and their key roles inME applications. Since the revival of thin-film ME heterostructures with giant ME coefficients\, a variety of multifunctional ME devices\, such as sensors\, inductors\, filters\, antennas etc. have been developed. Exciting progress has been made on novel ME materials and devices because of their high-performance ME coupling.\nIn this dissertation\, we will first show the properties of magnetostrictive (FeGaC and SmFe) and piezoelectric (ZnO)thin-film materials that are necessary for realizing strong ME coupling. A systematic investigation of the soft magnetism\, the change of modulus of elasticity with magnetization (delta-E effect)\, and microwave properties was carried out on FeGaC and SmFe thin films. We successfully developed the magnetostrictive FeGaC thin films with low coercive field of less than 1 Oe\, high saturation magnetization\, narrow ferromagnetic resonance (FMR) linewidth\, and an ultra-low Gilbert damping constant of 0.0027. A record high piezomagnetic coefficient of 9.71 ppm/Oe\, high saturation magnetostriction constant of 81.2 ppm\, and large delta-E effect of -120 GPa at 500 nm were achieved. ZnO films with high c-axis crystal orientation was also achieved by carefully optimizing the sputtering process parameters. These properties make them attractive materials for magnetoelectric and other voltage tunable RF/microwave device applications.\nAfter presenting the magnetostrictive and piezoelectric thin films and their static and dynamic properties\, we introduce the radio frequency (RF) ME microsystems. Mechanically driven antennas have been demonstrated to be the most effective method to miniaturize antennas compared to state-of-the-art compact antennas.The ME antennas based on a released magnetostrictive/piezoelectric heterostructure rely on electromechanical resonance instead of electromagnetic wave resonance\, which results in an antenna size as small as one-thousandth of an electromagnetic wavelength. Due to the strong ME coupling in thin-film ME heterostructures\, we proposed the ultra-compact MEMS ME antennas and improved their performance by using anchor designs\, array structure\, and SMR structure. These miniaturized robustME antennas can be implemented in numerous real-world applications such as internet of things\, wearable and bio-implantable devices\, smart phones\, wireless communication systems\, etc. The ME antennas\, with an overall dimension of 700 m×700 m (L×W)\, were designed to operate at a resonant frequency of 2 GHz and experimentally demonstrated a gain of -18.85 dBi. Furthermore\, we demonstrated highly sensitive integrated RF giant magnetoimpedance (GMI)sensors based on amplitude and phase sensitive mechanisms. The amplitude and phase magnetic noise levels were demonstrated to be 810pT /√Hz at 1000 Hz and 100pT /√Hz\, respectively.
URL:https://ece.northeastern.edu/event/ece-phd-dissertation-defense-xianfeng-liang/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210629T110000
DTEND;TZID=America/New_York:20210629T120000
DTSTAMP:20260426T130255
CREATED:20210623T211312Z
LAST-MODIFIED:20210623T211312Z
UID:5005-1624964400-1624968000@ece.northeastern.edu
SUMMARY:ECE PhD Dissertation Defense: Yaoshen Yuan
DESCRIPTION:PhD Dissertation Defense: Enhancing Monte Carlo Light Modeling Methods for the Development of Near-infrared Based Brain Techniques \nYaoshen Yuan \nLocation: Zoom Link \nAbstract: Studying light propagation in biological tissues is critical for developing biophotonics techniques and its applications. Monte Carlo (MC) method\, a stochastic solver for radiative transfer equation\, has been recognized as the gold standard for modeling light propagation in turbid media. However\, due to the stochastic nature of MC method\, millions even billions of photons are usually required to achieve accurate results using MC method\, leading to a long computational time even with the acceleration using graphical processing units (GPU).\nFurthermore\, due to the rapid advances in multi-scale optical imaging techniques such as optical coherence tomography (OCT) and multiphoton microscopy (MPM)\, there is an increasing need to model light propagation in extremely complex tissues such as vessel networks. The mesh-based Monte Carlo (MMC) is usually superior than the voxel-based MC method for such modeling since unlike grid-like voxels\, tetrahedral meshes can represent arbitrary shapes with curved boundaries. However\, the mesh density can be excessively high when the tissue structure is extremely complex\, resulting in high computational costs and memory demand.\nThe goal of this proposal is to focus on solving the challenges mentioned above. To tackle the first challenge\, we came up with a filtering approach with GPU acceleration to improve the signal-to-noise ratio (SNR) of the results while keeping the simulated photons low. The adaptive non-local means (ANLM) filter is selected to suppress the stochastic noise in MC results because 1) the filtering process on each voxel is mutually independent\, making it possible for parallel computing; 2) it has high performance in denoising and a strong capacity in edge-preserving.\nFor the second problem\, a novel method\, implicit mesh-based Monte Carlo (iMMC)\, was proposed to significantly reduce the mesh density. The iMMC utilizes the edge\, node and face of the tetrahedral mesh to model tissue structures with shapes of cylinder\, sphere and thin layer. The typical applications for edge\, node and face-based iMMC are vessel networks\, porous media and membranes\, respectively.\nLastly\, we applied MC simulations and aforementioned filter on segmented brain models derived from MRI neurodevelopmental atlas to estimate the light dosage for transcranial photobiomodulation (t-PBM)\, a technique for treating major depressive disorder using near infrared\, across lifespan. The iMMC simulation was also applied to evaluate the impact of human hair on the brain sensitivity for functional near-infrared spectroscopy (fNIRS). Furthermore\, a new approach that can improve the penetration depth in optical brain imaging as well as PBM is proposed. In this approach\, the possibility of placing light sources in head cavities is investigated using MC simulations. The preliminary results demonstrate a better performance in deep brain monitoring compared to the standard transcranial approach using 10-20 EEG positioning system.
URL:https://ece.northeastern.edu/event/ece-phd-dissertation-defense-yaoshen-yuan-2/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210622T140000
DTEND;TZID=America/New_York:20210622T150000
DTSTAMP:20260426T130255
CREATED:20210614T211638Z
LAST-MODIFIED:20210614T211638Z
UID:4986-1624370400-1624374000@ece.northeastern.edu
SUMMARY:PhD Dissertation Defense: Ala Tokhmpash
DESCRIPTION:PhD Dissertation Defense: Fractional Order Derivative in Circuits\, Systems\, and Signal Processing with Specific Application to Seizure Detection \nAla Tokhmpash \nLocation: Zoom Link \nAbstract: Epilepsy is a chronic brain disease that affects around 50 million people worldwide. This disease is characterized by recurrent seizures\, which are brief episodes of involuntary movement that may involve a part or the entire body and are sometimes accompanied by loss of consciousness. It is the third most common neurological disorder in the United States\, only after Alzheimer’s disease and stroke. Patients suffering from epilepsy\, a brain disorder\, can have more than one type of seizure. Seizure detection systems can be life-changing for patients with epileptic seizures. By accurately identifying the periods in which seizure occurrence has a higher chance of happening we can help epileptic patients live a more normal life. Prior works on automated seizure detection overwhelmingly either rely solely entirely on domain knowledge\, or instead use a black box deep learning model. This thesis aims to integrate machine learning techniques with available seizure detection methods to improve detection performance. In this process\, we take advantage of mathematical tools provided by fractional-order derivatives as well as fuzzy entropy concepts. Specifically\, 1) we show the effectiveness of fractional order methods (FOM) in representing signals with long-range dependencies 2) using case studies in control and power systems\, we further examine the performance of FOM in the presence of parameter uncertainty. 3) using two publicly available data sets of brain signals from multiple patients\, we develop a cohesive framework to leverage FOM for extracting features that can be then used by statistical learning methods. 4) following recent works in this field\, we generalize the notion of entropy to include the fractional-order case. Combined with the fuzzy sets describing the uncertainty in data\, we leverage fractional fuzzy entropy as a robust descriptor of the state of brain signals. Through these case studies\, we demonstrate a significant increase in performance accuracy compared to models that do not consider FOM.
URL:https://ece.northeastern.edu/event/phd-dissertation-defense-ala-tokhmpash/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210622T130000
DTEND;TZID=America/New_York:20210622T140000
DTSTAMP:20260426T130255
CREATED:20210617T174637Z
LAST-MODIFIED:20210617T174637Z
UID:4990-1624366800-1624370400@ece.northeastern.edu
SUMMARY:ECE PhD Dissertation Defense: Majid Sabbagh
DESCRIPTION:PhD Dissertation Defense: The Perils of Shared Computing: A Hardware Security Perspective \nMajid Sabbagh \nLocation: Microsoft Teams Link \nAbstract: Modern processors and hardware accelerators\, in the cloud or on the edge\, are capable of running multiple workloads from different users concurrently. Despite software techniques for security such as virtualization and containers\, a new attack surface is emerging that pertains to the hardware vulnerabilities of shared computing resources\, posing serious threats to shared computing. Fault attacks (FAs)\, Side-Channel Attacks (SCAs)\, and Transient-Execution Attacks (TEA) are three hardware-oriented attacks that target the system implementations. FAs aim to tamper the integrity of application execution through different fault injection methods\, to compromise the data or disrupt computation at run-time. SCAs exploit the information leakage of sensitive applications in physical parameters\, such as power consumption\, electromagnetic emanations\, and timing\, to breach the confidentiality of the application. TEAs exploit transient hardware operations such as speculative execution in Central Processing Units (CPUs) to tap on sensitive data temporarily and retrieve them from implicative microarchitectural states.\nIn this dissertation\, we investigate the three kinds of attacks that all exploit vulnerabilities due to shared computing. We first introduce a new non-invasive FA against Graphics Processing Units (GPUs)\, called overdrive fault attacks. We discover the security vulnerability of GPU’s voltage-frequency scaling (VFS) mechanism\, a common feature to balance power consumption and performance. An out-of-specification configuration of GPU voltage and frequency can be set by an adversary on the host CPU\, through the software interfaces to GPU’s power management units. This setting will cause timing violations for the computation and result in silent data corruptions (SDCs). We apply the overdrive fault attacks on two common victim applications. One is cryptographic applications accelerated by GPU. We launch a differential fault analysis (DFA) attack on an AES kernel running on an AMD RX 580 GPU and successfully recover the secret key. The other victim is convolutional neural network (CNN) inference. We thoroughly characterize fault injections and propagation in a CNN on a GPU and analyze the controllability of the attack. We successfully launch an end-to-end misclassification attack during CNN inferences with careful timing control.\nWe then evaluate a timing side-channel attack called Prime+Probe attack on CPUs and propose a Side-Channel Attack DEtection Tool (SCADET). SCADET is a methodology and a tool that operates on an x86 program’s binary. It records and analyzes the program’s memory accesses using dynamic binary instrumentation by running the program in a controlled environment to accurately identify the malicious access patterns demonstrated by the Prime+Probe attack.\nFinally\, we introduce an efficient hardware-level taint-tracking defense against the most prominent TEAs\, the speculative execution attacks. We take a secure-by-design approach and propose a mechanism called Secure Speculative Execution via RISC-V Open Hardware Design (SSE-RV)\, based on the latest Berkeley Out-of-Order Machine (SonicBOOM). We prototype our SSE-RV processor on an FPGA running a Linux operating system. Our results show that we can protect against Spectre-v1\, v2\, and v5. Our defense scheme is general and can be extended to protect against other transient execution attacks.
URL:https://ece.northeastern.edu/event/ece-phd-dissertation-defense-majid-sabbagh/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210618T090000
DTEND;TZID=America/New_York:20210618T100000
DTSTAMP:20260426T130255
CREATED:20210614T181136Z
LAST-MODIFIED:20210614T181136Z
UID:4983-1624006800-1624010400@ece.northeastern.edu
SUMMARY:PhD Dissertation Defense: Trinayan Baruah
DESCRIPTION:PhD Dissertation Defense: Improving the Virtual Memory Efficiency of GPUs \nTrinayan Baruah \nLocation: Zoom Link \nAbstract: GPUs have been adopted widely based their ability to exploit data-level parallelism found in modern-day applications\, ranging from high performance computing to machine learning. This widespread adoption has\, in part\, been accelerated by the development of more intuitive high-level programming languages\, efficient runtimes and drivers\, and easier mechanisms to manage data movement. Modern day GPUs and multi-GPU systems utilize virtual memory systems\, enabling programmers to access large address spaces that are beyond the physical memory limits a GPU. There mechanisms have built in mechanisms for memory translation\, sparing the programmer from having to reason about complex data-movement operations. Virtual memory support on a GPU includes both hardware and software support. At the hardware level\, translation lookaside buffers~(TLBs) are used to cache translations close to the compute units. At the software level\, the programming model supports a unified memory model which automates the movement of pages across multiple devices in a a system. Despite the improvements in programmability\, due to inefficiencies existing virtual memory management mechanisms\, including TLB management and page migration policies\, the performance obtained on today’s GPUs is sub-optimal.\nIn this dissertation\, we first identify the key challenges in virtual memory support for GPUs today. We then propose mechanisms to reduce the bottlenecks arising from virtual memory support at both a hardware level and at the runtime level. This allows GPUs to fully enjoy the benefits of virtual memory\, while ensuring high performance. We also develop simulation tools that enable researchers to explore new and novel virtual memory features in future single GPU and multi-GPU systems.\nTo enhance hardware support for virtual memory on a GPU\, we explore a mechanism that enables prefetching of page-table entries into the GPUs TLBs\, thereby reducing the number of TLB misses and improving performance. We also leverage the fact that many page-table entries can be shared across different GPU cores. We design a low-cost interconnect that enables sharing of page-table entries across the GPU cores. To improve the performance of unified memory on multi-GPU systems\, we propose a hardware/software mechanism that monitors accesses to each page\, and uses this information when making page-migration decisions. We also propose mechanisms to reduce the cost of TLB shootdowns on the GPU during page-migration in NUMA multi-GPU systems.
URL:https://ece.northeastern.edu/event/phd-dissertation-defense-trinayan-baruah/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210614T090000
DTEND;TZID=America/New_York:20210614T100000
DTSTAMP:20260426T130255
CREATED:20210602T181602Z
LAST-MODIFIED:20210602T181602Z
UID:4966-1623661200-1623664800@ece.northeastern.edu
SUMMARY:Robotics at Northeastern: A personal view of Autonomy Underwater\, On Land and with Aerial Systems
DESCRIPTION:This session will cover interconnected nature of autonomous vehicles on land\, underwater\, and in the air. Using examples from research at Northeastern\, it explores the interconnections of systems design\, application-driven robotics\, and the fundamental algorithms associated with simultaneous localization and mapping (SLAM)\, machine learning (ML)\, and computer vision and imaging as it applies to a variety of scientific and commercial settings. \nThis webinar will be lead by ECE Faculty Member Dr. Hanumant Singh. Event logistics are below: \nDate: June 14\, 2021 \nTime: 9:00am EST \nRegistration \n 
URL:https://ece.northeastern.edu/event/robotics-at-northeastern-a-personal-view-of-autonomy-underwater-on-land-and-with-aerial-systems/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210611T103000
DTEND;TZID=America/New_York:20210611T113000
DTSTAMP:20260426T130255
CREATED:20210609T174749Z
LAST-MODIFIED:20210609T174749Z
UID:4976-1623407400-1623411000@ece.northeastern.edu
SUMMARY:PhD Proposal Review: Rashida Nayeem
DESCRIPTION:PhD Proposal Review: Human Strategies in the Control of Complex Objects: A Task-Dynamic Approach with Clinical Applications \nRashida Nayeem \nLocation: Zoom Link \nPasscode: 247537 \nAbstract: Functional interaction with objects – tool use – is essential in daily living and is regarded as the foundation of our evolutionary advantage. Humans effortlessly interact with a variety of objects\, including those with complex internal dynamics. Even the simple action of picking up a cup of coffee to drink is a mechanically intricate process: the hand applies a force to the cup\, and indirectly to the liquid\, which exerts forces back on the hand. Reacting to and mitigating these dynamics in real time is difficult due to long sensorimotor delays and ubiquitous noise in the human sensorimotor system. Hence\, prediction is necessary to preempt undesired ‘sloshing’. But prediction of this nonlinear and potentially chaotic object is extremely difficult. Hence\, this research tests the hypothesis that humans learn to control the object to make dynamics simpler––or more predictable. Inspired by the task of transporting a ‘cup of coffee\,’ a series of experiments use a virtual and a real testbed that model the task as a cup-and-ball system. In all experiments\, human subjects move the cup with a rolling ball inside. Aim-1 investigates the effect of linearization on human control strategies in a 2D virtual task. Aim-2 examines if humans exploit initial conditions to facilitate predictable dynamics in the 2D virtual task. This is tested in subjects that are provided with either full sensory information\, or when deprived of visual or haptic information. Aim-3 examines how subjects explore and prepare the cup-and-ball in the same 2D task and a novel 3D virtual task introducing planar cup movements. The question is how subjects explore and transport objects that have different dynamic properties\, either unknown or indicated by the visual cues. The analysis adopts a task-dynamic approach that affords principled hypothesis-testing by parsing the complex dynamics into execution and result variables\, with minimal assumptions about the human controller. Aim-4 takes the insights from this basic research to a clinical context\, testing patients with stroke in this functional task. A real version of the cup-and-ball task was created to quantitatively assess severity and recovery of motor impairment in patients after stroke. Using the same analysis methods\, the objective is to sensitively assess impairment in the context of a functional skill.
URL:https://ece.northeastern.edu/event/phd-proposal-review-rashida-nayeem/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210610T153000
DTEND;TZID=America/New_York:20210610T163000
DTSTAMP:20260426T130255
CREATED:20210609T174644Z
LAST-MODIFIED:20210609T174644Z
UID:4974-1623339000-1623342600@ece.northeastern.edu
SUMMARY:PhD Proposal Review: Bernard Herrera
DESCRIPTION:PhD Proposal Review: Ferroelectric Micromachined Ultrasonic Transducers for Intra-body and In-memory Sensing Applications \nBernard Herrera \nLocation: Teams \nAbstract: Piezoelectric Micromachined Ultrasonic Transducers (PMUTs) are Micro Electro-Mechanical Systems (MEMS) devices that have become an established technology in applications such as range-finding\, fingerprint sensing and imaging due to their capability of ultrasonic transduction in a miniaturized footprint\, easily amenable to create large arrays. However\, their application space still remains quite open. PMUTs are well fitted to applications in liquid media\, such as implantable and underwater devices\, due to their inherent acoustic matching and wide bandwidth. Thus\, in the first part of this proposal\, we explore novel applications such as PMUT-based intra-body networking\, power transfer\, source localization\, wide-band matching and duplexing.\nAluminum Nitride (AlN) has been the material of choice for our PMUTs due to its biocompatibility and possibility of single-chip integration with supporting CMOS circuitry. Scandium doping of AlN thin films has recently been demonstrated to increase piezoelectric coupling coefficients while introducing ferroelectric properties in the material. However\, a simultaneous use of both capabilities has not been demonstrated in the state-of-the-art. The ability of having distinct ferroelectric states\, that alter the mechanical performance of the devices\, allows for in-memory sensing and actuation features and provides the building blocks for neuromorphic signal processing capabilities. The second part of the proposal explores the AlScN material integration into novel Ferroelectric Micromachined Ultrasonic Transducers (FMUTs) and their emerging application space.
URL:https://ece.northeastern.edu/event/phd-proposal-review-bernard-herrera/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210610T130000
DTEND;TZID=America/New_York:20210610T140000
DTSTAMP:20260426T130255
CREATED:20210510T185954Z
LAST-MODIFIED:20210510T185954Z
UID:4931-1623330000-1623333600@ece.northeastern.edu
SUMMARY:ECE PhD Proposal Review: Andac Demir
DESCRIPTION:PhD Proposal Review: Automated Bayesian Network Exploration for Nuisance-Robust Inference \nAndac Demir \nLocation: Zoom Link \nAbstract: A fundamental challenge in the analysis of physiological signals is learning useful features that are robust to nuisance factors e.g.\, inter-subject & inter-session variability\, and achieve the highest nuisance-invariant classification performance. Towards resolving this problem\, we introduce 2 frameworks: AutoBayes\, which is an AutoML approach to conduct neural architecture search for research prototyping\, and a GNN based framework: EEG-GNN.\nThe ultimate goal of the AutoBayes framework is to identify the conditional relationship between a physiological dataset\, associated task labels\, nuisance variations and potential latent variables in order to robustly infer the task labels invariant of nuisance factors. Nuisance factors in the case of physiological datasets could be variations in subjects or sessions\, but we only focus on subject variations in the experiments. AutoBayes enumerates all plausible Bayesian networks between data\, labels\, nuisance variations and potential latent variables\, detects and prunes the unnecessary edges according to Bayes-Ball Algorithm\, and then trains the resulting DNN architectures for different hyperparameter configurations in an adversarial/non-adversarial or a variational/non-variational setting to achieve the highest validation performance. Instead of hyperparameter tuning for model optimization\, AutoBayes concentrates on the architecture search of plausible Bayesian networks\, and achieves state-of-the-art performance across several physiological datasets. Furthermore\, we ensemble several Bayesian networks by stacking their posterior probability vectors in a higher level learning space\, train a shallow MLP as a meta learner\, and measure the task and nuisance classification performance on a hold-out dataset. We observe that exploration of different inference Bayesian networks has a significant benefit in improving the robustness of the machine learning pipeline\, and the parallel activity of vast assemblies of different Bayesian network models significantly reduces variation across subjects in the cross-validation setting.\nIn the second part of the proposal\, we benchmark the performance of EEG-GNN against the AutoBayes framework. CNN’s have been frequently used to extract subject-invariant features from EEG for classification tasks\, but this approach holds the underlying assumption that electrodes are equidistant analogous to pixels of an image and hence fails to explore/exploit the complex functional neural connectivity between different electrode sites. We overcome this limitation by tailoring the concepts of convolution and pooling applied to 2D grid-like inputs for the functional network of electrode sites. Furthermore\, we develop various GNN models that project electrodes onto the nodes of a graph\, where the node features are represented as EEG channel samples collected over a trial\, and nodes can be connected by weighted/unweighted edges according to a flexible policy formulated by a neuroscientist. The empirical evaluations show that our proposed GNN-based framework outperforms standard CNN classifiers across ErrP and RSVP datasets\, as well as allowing neuroscientific interpretability and explainability to deep learning methods tailored to EEG related classification problems. Another practical advantage of our GNN-based framework is that it can also be used in EEG channel selection\, which is critical for reducing computational cost\, and designing portable EEG headsets.
URL:https://ece.northeastern.edu/event/ece-phd-proposal-review-andac-demir/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210608T100000
DTEND;TZID=America/New_York:20210608T110000
DTSTAMP:20260426T130255
CREATED:20210602T213048Z
LAST-MODIFIED:20210602T213048Z
UID:4968-1623146400-1623150000@ece.northeastern.edu
SUMMARY:ECE PhD Dissertation Defense: Lorenzo Bertizzolo
DESCRIPTION:PhD Dissertation Defense: Software-Defined Wireless Networking for 5G and Beyond: From Indoor Cells to Non-Terrestrial UAV Networks \nLorenzo Bertizzolo \nLocation: Microsoft Teams \nAbstract: While Software-defined Networking (SDN) is a consolidated and widely adopted concept in fixed infrastructure\, its adoption to the wireless domain has been limited by some fundamental challenges. Unlike traditional fixed infrastructure\, which relies entirely on fiber communications\, wireless networks suffer from unpredictable access and backhaul operations. Therefore\, they ask for new architectural solutions to implement SDN. A prerequisite to implementing SDN for wireless is to adopt programmable radio hardware and software-based implementation of the wireless protocol stack. Then\, it is necessary to adopt control APIs that expose the wireless protocols’ internal operations to external control. Finally\, to bridge the gap between SDN theory and implementation\, this thesis proposes a series of architectural solutions that provide the communication infrastructure and the architectural control innovations necessary to implement SDN control and optimization on real wireless systems. Specifically\, this thesis proposes two different architectural solutions. For fixed Radio Access Networks (RANs) that benefit from a low-latency and reliable backhaul infrastructure\, we propose to implement a centralized control approach similar to SDN for fixed infrastructure. Here a central controller collects network state information of a distributed RAN\, solves a network control problem\, and distributes the solutions to the individual wireless nodes. On the other hand\, for infrastrcuture-less RANs\, which implement both the access and the backhaul in wireless\, we propose a new architectural solution that moves the control logic to the edge of the networks\, to the very wireless nodes.
URL:https://ece.northeastern.edu/event/ece-phd-dissertation-defense-lorenzo-bertizzolo/
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210603T150000
DTEND;TZID=America/New_York:20210603T160000
DTSTAMP:20260426T130255
CREATED:20210602T213220Z
LAST-MODIFIED:20210602T213220Z
UID:4970-1622732400-1622736000@ece.northeastern.edu
SUMMARY:ECE PhD Dissertation Defense: Yumin Liu
DESCRIPTION:PhD Dissertation Defense: Learning from Spatio-Temporal Data with Applications in Climate Science \nYumin Liu \nLocation: Zoom Link \nAbstract: Climate change is one of the major challenges to human beings and many other species in our time. In the recent decade\, the number of disasters related to climate change such as wildfires\, storms\, floods and droughts are increasing\, and the casualty and economic losses caused by them are larger compared to those of decades ago. This calls for better and efficient ways to predict climate change in order to better prepare and reduce losses. Predicting climate change involves using historical observational data and model simulated data\, both of which usually involve multiple locations and timestamps and are spatio-temporal. With the rapid development and progress of machine learning\, these methods have achieved several impactful contributions in many domains; we would like to translate its impact to climate science.\nIn this thesis we address several problems in climate science. This challenging complex domain enable us to develop\, innovate\, adapt\, and advance machine learning in the following ways. 1) We develop a multi-task learning method to estimate relationships between tasks and learn the basis tasks in different locations especially for nearby locations which may share similar climate patterns. This method assumes that the weights of an observed task is a linear combination of several latent basis tasks and that the task relationships can be learnt by imposing a regularized precision matrix. 2) We propose a nonparameteric mixture of sparse linear regression models to cluster and identify important climate models for prediction. This model incorporates Dirichlet Process (DP) to automatically determine the number of clusters\, imposes Markov Random Field (MRF) constraints to guarantee spatio-temporal smoothness\, and selects a subset of global climate models (GCMs) that are useful for prediction within each spatio-temporal cluster with a spike-and-slab prior. We derive an effective Gibbs sampling method for this model. 3) We adapt image super resolution methods to climate downscaling — increasing spatial resolution for climate variables for local impact analysis. The proposed method is called YNet which is a novel deep convolutional neural network (CNN) with skip connections and fusion capabilities to perform downscaling for climate variables on multiple GCMs directly rather than on reanalysis data. 4) We use saliency map method to discover dependencies among climate variables. We propose the concept of cyclical saliency map (Cyclic-SM) which are meaningful in climate context and more robust to noise as compared to ordinary saliency maps. We show that Cyclic-SMs can reveal relevant spatial regions for prediction. We demonstrate the effectiveness of this method in climate downscaling\, ENSO index prediction and river flow prediction.
URL:https://ece.northeastern.edu/event/ece-phd-dissertation-defense-yumin-liu/
END:VEVENT
END:VCALENDAR