PhD Colloquia

Spring 2017

Spring Seminars will be held in three different places:
Towne 337 -- 220 S. 33rd Street
Singh 035 -- 3205 Walnut Street
PERCH 301 -- 3401 Greys Ferry Ave. View on Google Maps. Penn Transit Services offers an on-demand shuttle service to Pennovation Works, call 215.898.RIDE (7433).

Wednesday, February 22
Gopinath Danda
"Two-Dimensional Nanopore Sensors for Molecular Detection and Analysis"
Towne 337

Read the Abstract and Bio

Abstract: Solid-state nanopore sensors can detect and analyze molecules by monitoring the change in ionic current through a nanopore as the molecule translocates through, leading to applications like DNA sequencing and nanoparticle charge analysis. As is true with any new technology, there are existing challenges associated with these sensors, including low signal-to-noise ratio (SNR), conventional low-throughput transmission electron microscope (TEM) based fabrication methods and fast DNA translocation speeds (hindering sequencing). In this talk, I will discuss how atomically thin two-dimensional membranes, like graphene and transition metal dichalcogenides (TMDs), can help tackle some of these problems while also introducing unique difficulties and functionalities.

Bio: Gopinath Danda received his B.Tech. degree in Electronics and Communication Engineering from KIIT University, Bhubaneswar, India, in 2012, before finishing his M.S. in Nanotechnology from University of Pennsylvania, Philadelphia, US, in 2014. He is currently working towards his Ph.D. degree in the Department of Electrical and Systems Engineering, University of Pennsylvania, under the supervision of Prof. Marija Drndić in the Physics & Astronomy Department. His research interest lies in two-dimensional material based nanodevices, with current focus in nanopore sensors for molecular detection and desalination, and investigation of optical and electronic properties of nanostructures.

Wednesday, March 1
Mehrdad Pourfathi
"Hyperpolarized Magnetic Resonance Imaging and Spectroscopy"
Singh 035

Read the Abstract and Bio

Abstract: Magnetic resonance imaging (MRI) is a noninvasive diagnostic1H) nucleus, other nuclei (13C, 3He, 129Xe, etc.) can also be used for the in-vivo imaging of functional and metabolic activity. However, the utility of these alternative nuclei is hindered by the inherently lower signal-to-noise due to the significantly lower physiological concentrations as well as lower gyromagnetic ratio. Hyperpolarization is a mechanism that dramatically increases the signal (polarization) of these nuclei, resulting in enhanced sensitivity of MRI by a factor of 10,000-100,000. However, the hyperpolarization lifetime is limited by irreversible relaxation mechanisms at the molecular level. As such, the development of special imaging techniques is required for rapid and efficient data acquisition. Despite these challenges, hyperpolarized MRI has shown great promise as a research and preclinical tool to reveal information about physiology and biochemistry since its emergence twenty years ago. The objective of this presentation is to provide a brief introduction to hyperpolarized MRI and the ongoing research at the University of Pennsylvania.

Bio: Mehrdad Pourfathi is currently a Ph.D. student of Electrical and Systems Engineering at the University of Pennsylvania. He is currently a member of the Functional and Metabolic Imaging group at the Radiology Department at the Perelman School of Medicine, University of Pennsylvania, under the supervision of Dr. Rahim Rizi. His primary research focuses on the applications of hyperpolarized MRI to molecular imaging of the lungs. More specifically, his work involves: (1) investigating sample preparation techniques for hyperpolarization of xenon-129 and carbon-13-based biomarkers using dynamic nuclear polarization (DNP), (2) developing rapid MRI pulse sequences for metabolic and pH imaging of the lungs (3) studying the alterations in lung metabolism and pH in lung inflammatory diseases.


March 8 - SPRING BREAK - No Seminar

Wednesday, March 15
Min Wen
Learning with the Task in Mind: Provably Safe Reinforcement Learning

Read the Abstract and Bio

Abstract: With increasing success in AI and robotics, reinforcement learning (RL) is recognized as an effective tool to implement complex tasks that are encoded as optimization objectives with carefully designed reward functions. However, reward functions as task descriptions, suffer from several noticeable drawbacks: First, it is not a trivial task to design reward functions such that an optimal policy will successfully implement the designated task. Second, if global optimum is not achievable, suboptimality in optimization objective is usually not a good measure of the degradation of task implementation quality. Last but not least, reward functions can provide no safety guarantee during exploration in unknown environments, which is a key concern in safety-critical systems.

To alleviate such problems, we make use of probabilistic temporal logic to encode high-level task requirements that should never be violated (even during learning), which is a lot easier for human to specify and understand. In this talk, I will introduce the problem of provably safe reinforcement learning, in which we do reinforcement learning tasks while keeping track of high-level task requirements. We propose reinforcement learning algorithms in the setting of stochastic games to learn near-optimal policies that are safe during learning and guarantted to implement tasks with high probability. We also address a problem of inverse reinforcement learning with high-level task requirements available as side information, and show the necessity of high-level information to reliably learn to implement our task from demonstrations.

Bio: Min Wen received the B.S. degree in Automation from Zhejiang University, Hangzhou, China in 2013. She is currently pursuing the Ph.D. degree in the Electrical and Systems Engineering program at the University of Pennsylvania. Her research interests include machine learning, optimal control and formal methods, with special focus on learning based control with formal correctness guarantees.

Wednesday, March 22
Jinwook Huh
Manipulation and Grasping with Learning High-Dimensional Mixture Models in Constrained Sampling-Based Planning
Towne 337

Read the Abstract and Bio

Abstract: In recent years, there has been growing interest in sampling-based motion planners for grasping and manipulation. However, sampling based approaches have a critical problem that they spend most of their computational time on checking for collisions, and they are difficult to involve constraints on end-effector pose for grasping and manipulation with the high dimensional manipulator. In this talk, we present a new approach for fast collision detection in high dimensional configuration spaces based on the Gaussian Mixture Models (GMMs) for Rapidly-exploring Random Trees (RRT) motion planning. It is also able to adapt to environmental change by learning models of collision and collision-free regions in configuration spaces in an online manner. In addition, we will discuss a novel constrained sampling-based motion planning method for grasping and transport tasks with a redundant robotic manipulator, which allows the best grasp configuration and approach direction to be automatically determined. The method also introduces a parameterized intermediate pose that is optimized to determine the approach direction, increasing robustness under sensor uncertainty and execution errors. Our approach also considers transporting the grasped object to the desired target position using the RRT algorithm that incorporates soft constraints via appropriate cost penalties.

Bio: Jinwook Huh received the BS and MS in mechanical engineering from POSTECH, Korea in 2004 and 2006, respectively. He worked for the Agency for Defense Development, Korea from 2006 to 2013. He is currently a Ph.D. student in the Department of Electrical and Systems Engineering and the GRASP Laboratory. His research interests include motion planning based on the machine learning and autonomous vehicle system.

Wednesday, March 29
Achin Jain
Bridging the Gap Between Machine Learning and Predictive Control
Singh 035

Read the Abstract and Bio

Abstract: Machine learning and control theory are two foundational but disjoint communities. Machine learning requires data to generate models, and control systems require models to provide stability and performance guarantees to plant operations. Machine learning is widely used for regression or classification, but thus far the data-driven models have not been suitable for closed-loop control. In many applications such as building and process control, model identification using first principles becomes cost and time prohibitive with increasing scale and complexities, thus limiting the use of model-based controllers like MPC. The challenge now is to develop new learning algorithms that are also suitable for control.

In this talk, I will present our framework for Data Predictive Control (DPC). DPC takes historical data from the sensors that are already installed to learn control-oriented models, thus bypassing the need for expensive white-box modeling. In particular, I will discuss novel algorithms based on regression trees and random forests for receding horizon control. On the applications side, I will talk about DPC for demand response and building climate control.

Bio: Achin Jain is a doctoral candidate in Electrical and Systems Engineering at the University of Pennsylvania. He received M.Sc. in Robotics, Systems and Control from ETH Zurich, Switzerland in 2015 and B.Tech. in Mechanical Engineering from the Indian Institute of Technology Delhi, India in 2012. He worked at Daimler AG and ABB Corporate Research in 2013-14. His research interests include machine learning, control theory, optimization and statistics applied to cyber-physical systems. His current research focuses on control with data-driven models.

Image result for matthew o'kelly

Wednesday, April 5
Matthew O'Kelly
Check Yourself, Before You Wreck Yourself: Developing Trustworthy Autonomous Vehicles

Read the Abstract and Bio

Abstract: The goal of the APEX (Autonomous Vehicle Plan Execution and Verification) Framework is to test and verify the decision controllers of autonomous vehicles (AVs). If AVs are to be considered drivers, as the NHTSA tepidly recommended to AV manufacturers in a recent letter, we must derive new solutions to bound the risk that the vehicle’s software poses to society.

Complete AV safety requires that the consequences of the software agent’s actions are considered over the set of all possible configurations. The APEX approach begins with a scenario description language which provides a means of instantiating autonomous agents within a variety of multi-objective scenarios. To this end, the talk will address the models used within the APEX agent library, and the structure of a scenario in the context of the hybrid systems formalism.

Given an APEX scenario description, the tool automatically generates executable models, which are proved to be safe or unsafe with respect to the specification governing the system. We consider both the use of delta-decision procedures to generate formal proofs, and enhancements to such algorithms using robustness metrics derived from MCMC simulations. In order to provide actionable feedback, we demonstrate the concretization and visualization of counterexamples.

The second half of the talk addresses the scalability of such methods via PAC learning of the minimum robustness metric over the entire configuration space, the introduction of continuous approximations of optimal controllers, and photorealistic simulations which leverage existing rendering engines.

Bio: Matthew O’Kelly is a NSF GRFP fellow and Ph.D candidate in Electrical and Systems Engineering at the University of Pennsylvania. In the summer of 2015, he was an NSF EAPSI fellow at Nagoya University’s Parallel and Distributed Systems Lab where he contributed the motion planning algorithms for the Autoware open source autonomous vehicle operating system. Prior to joining the University of Pennsylvania, he received a B.S. and M.S. in Mechanical Engineering from The Ohio State University. His research interests span cyber-physical systems under the broad goal of embedding trustworthy, personalized autonomy in everyday objects.

Wednesday, April 12
Yash V. Pant
Co-design of Computation and Control for Autonomous Systems: A Robust Predictive Control Approach
Towne 337

Read the Abstract and Bio

Abstract: Control software of autonomous robots has stringent real-time requirements that must be met to achieve the control objectives. One source of variability in the performance of the closed loop system is the execution time and accuracy of the state estimator, which is generally perception based. The first part of the talk will define a framework for co-designing anytime estimation and control algorithms, in a manner that accounts for implementation issues like delays and inaccuracies. With a focus on stabilization and trajectory tracking, the predictive control algorithm presented will also provide formal guarantees on stability and feasibility of the system. The second part of the talk will focus on control of systems with complex objectives that go beyond the traditional tracking and stability, with these complex spatio-temporal requirements represented as Metric Temporal Logic formulae. To meet these requirements, a predictive control algorithm will be presented that avoids the pitfalls of Mixed Integer Linear Programming and stochastic heuristics based approaches which are state of the art for such a problem.

Bio: Yash Vardhan Pant is a PhD Student, advised by Prof. Rahul Mangharam, at the Department of Electrical and Systems Engineering at the University of Pennsylvania. He received his MSE degree in Electrical Engineering from the University of Pennsylvania in 2012, and has been a recepient of the Richard K. Dentel memorial award for research in Urban transportation from Penn.

Wednesday, April 19
Wenxiang Chen
Mechanically Tunable Surface Lattice Resonances for Au Gratings on Elastomeric Substrates
Singh 035

Read the Abstract and Bio

Abstract: Metamaterials commonly have a fixed operational wavelength or can only operate over a limited bandwidth. Actively tunable metamaterials, which have large frequency or amplitude tunability, are important in applications such as sensors, metalenses, terahertz switches and modulators. We fabricated Au grating structures on elastomeric polydimethylsiloxane (PDMS) substrates, and studied different designs to mechanically enhance the tunability of the grating pitch. Finite-difference time-domain simulations predict narrow bandwidth resonance peaks and large frequency variation with strain. Optical measurements are being carried out.

Bio: Wenxiang Chen is a doctoral student in Department of Electrical and Systems Engineering at the University of Pennsylvania from 2011. He received his bachelor degree in Applied Physics in University of Science and Technology of China in 2011. His current work focuses on developing nanocrystal building blocks for plamonic devices, such as nano-antennas, optical poalrizers and plasmonic solar cells, and fabrication of plamsonic structures on flexible substrates.

Wednesday, April 26
Kuk Jang
Towards Computer-Aided Clinical Trials of Implantable Cardioverter Defibrillators

Read the Abstract and Bio

Abstract: Medical devices like the Implantable Cardioverter Defibrillator (ICD) are life-critical systems. Malfunctions of the device can cause serious injury or death of the patient. The current gold standard for proving the efficacy and safety of high-risk medical devices such as the ICD is through the results of a clinical trial (CT). CTs can run for several years, cost millions of dollars, and pose an inherent risk to the patients by exposing them to an unproven device.

We have developed a framework which allows the use of large model-based synthetic groups of patients and device models to improve the planning and execution of a CT so as to increase the chances of a successful trial. We demonstrate the utility of the framework by applying it to RIGHT, a clinical trial which compared two ICDs from different manufacturers. We show that the conclusions from our framework corroborate the results of the RIGHT trial and how our in-silico framework can provide additional insight into device performance that would not be possible in-vivo .

In addition, some of ongoing efforts to develop generative models of the synthetic group of patients using data-driven methods will be presented in the talk. In particular, we apply Long Short-Term Memory (LSTM) Networks to model electrogram signals and generate synthetic electrograms which can be used to test ICD discrimination algorithms.

Bio: Kuk Jin Jang is a PhD candidate in the department of Electrical and Systems Engineering at the University of Pennsylvania, advised by Prof. Rahul Mangharam.
He received an M.Eng degree from Princeton University and a Sc.B from Brown University. From 2011 to 2013, he worked with the Korea Electronics Technology Institute on interface systems for biochemical sensors. His current research interests include modeling of physiological processes, design and verification of medical cyber physical systems.

Wednesday, May 3
Ximing Chen
Minimal Network Perturbation for Structural Controllability
Towne 337, 12:15p

Read the Abstract and Bio

Abstract: We address the problem of optimally modifying the topology of a directed dynamical network to ensure structural controllability. More precisely, given the structure of a directed dynamical network (i.e., an existing networked infrastructure), we propose a framework to find the minimum number of directed edges that need to be added to the network topology in order to render a structurally controllable system. Our main contribution is twofold: (i) we provide a full characterization of all optimal network modifications, and (ii) we propose an algorithm able to find an optimal solution in polynomial time. We illustrate the validity of our algorithm via numerical simulations in random networked systems.

Bio: Ximing Chen received his Bachelor degree in Electronic Engineering from the Hong Kong University of Science and Technology, Hong Kong, China, in 2013. Since then, he has been working towards the Ph.D. degree in the Department of Electrical and System Engineering, University of Pennsylvania. His research interests include analysis and control of networked systems, with applications in social network, epidemic spreading and transportation.

Wednesday, May 10
Didi She
Immobilized Electrolyte Biodegradable Batteries for Implantable MEMS
Singh 035

Read the Abstract and Bio

Abstract: Recent progress in biodegradable medical devices motivates the development of
similarly biodegradable batteries to support self-powered transient systems. Liquid electrolyte
volume in microfabricated biodegradable batteries is a key driver in the lifetime and overall size of electrochemical cells. Harnessing liquid from the body to act as an electrolyte is therefore desirable; however, for stable operation, maintaining a constant environment inside the electrolytic cell is required even in the presence of changing body conditions. In this talk we present a biodegradable battery featuring a solid electrolyte of sodium chloride (NaCl) and polycaprolactone (PCL). The large excess of ionic material suspended in the PCL acts to hold intracellular conditions constant in the presence of varying external aqueous conditions, thereby achieving reduced cell size compared to the state of the art while maintaining performance stability.

Bio: Didi She received the B.S. degree in Electrical Engineering from Southeast University,
Nanjing, China in 2011, and M.S. degree in Microelectronics from Peking University, Beijing,
China in 2014. She is currently a Ph.D. student in the Department of Electrical and Systems
Engineering and the MSMA Laboratory. Her research interests include biocompatible and
biodegradable medical devices for implantable MEMS.

Fall 2016

Seminars will be held in Towne 337 on Wednesdays at 12:00 PM unless otherwise specified.

September 7
Wenxiang Chen
Hydrogel-covered Au Nanorod array optical humidity sensor for agriculture use
Click here to view Wenxiang Chen's seminar

Read the Abstract and Bio

Abstract: Innovation of technology has been a promising way to address challenges in agriculture. Here, we report a hydrogel-covered Au nanorod array optical humidity sensor for agriculture use. It works by transducing refractive index changes of the hydrogel upon exposure to moisture into spectral shifts of the Au nanorod array localized surface plasmon resonance (LSPR). We tune the sensor operation wavelength from 1205 nm to 2413 nm by tailoring the nanorod length and the recovery time from 1.9min to 27 h by controlling the hydrogel thickness. The sensitivity is up to 0.72 nm / % relative humidity (RH) in 23.8% to 86.8% RH range. At mist environment the sensor resonance moves according to the visibility changes. The strong, tunable, and environmentally sensitive LSPR of the Au nanorods make them ideal platforms upon which to design optical sensors for remote sensing.

Bio: Wenxiang Chen is a doctoral student in Department of Electrical and Systems Engineering at the University of Pennsylvania from 2011. He received his bachelor degree in Applied Physics in University of Science and Technology of China in 2011. His current work focuses on developing nanocrystal building blocks for plasmonic devices, such as nano-antennas, optical polarizers and plasmonic solar cells, and fabrication of plasmonic structures on flexible substrates.

September 14
Amin Rahimian
Bayesian Heuristics for Group Decisions
Click here to view Amin Rahimian's seminar

Read the Abstract and Bio

Abstract: We propose a model of inference and heuristic decision making in groups that is rooted in the Bayes rule but avoids the complexities of rational inference in partially observed environments with incomplete information. According to our model, the group members behave rationally at the initiation of their interactions with other group members; however, in the ensuing decision epochs they rely on a heuristic that replicates their experiences from the first stage. Subsequently, the agents use their time-one Bayesian update and repeat it for all future time-steps; hence, updating their actions using a so-called Bayesian heuristic. This model avoids the complexities of fully rational inference and also provides a behavioral and normative foundation for non-Bayesian updating. It is also consistent with a dual-process psychological theory of decision making, where a controlled (conscious/slow) system develops the Bayesian heuristic at the beginning, and an automatic (unconscious/fast) system takes over the task of heuristic decision making in the sequel.
We specialize this model to a group decision scenario where private observations are received at the beginning, and agents aim to take the best action given the aggregate observations of all group members. We present the implications of the choices of signal structure and action space for such agents. We show that for a wide class of distributions from the exponential family the Bayesian heuristics take the form of an affine update in the self and neighboring actions. Furthermore, if the priors are non-informative (and possibly improper), then these action updates become a linear combination. We investigate the requirements on the modeling parameters for the action updates to constitute a convex combination as in the DeGroot model. The results reveal the nature of assumptions that are implicit in the DeGroot updating and highlights the fragility and restrictions of such assumptions; in particular, we show that for a linear action update to constitute a convex combination the precision or accuracy of private observations should be balanced among all neighboring agents, requiring a notion of social harmony or homogeneity in their observational abilities. Following the DeGroot model, agents reach a consensus asymptotically. We derive the requirements on the signal structure and network topology such that the consensus action aggregates information efficiently. This involves additional restrictions on the signal likelihoods and network structure; in particular, among all connected and undirected social networks only complete graphs can lead to efficient consensus.

We next shift attention to a finite state model, in which agents take actions over the probability simplex; thus revealing their beliefs to each other. We show that the Bayesian heuristics in this case prescribe a log-linear update rule, where each agent's belief is set proportionally to the product of her own and neighboring beliefs. We analyze the evolution of beliefs under this rule and show that agents reach a consensus. The consensus belief is supported over the maximizers of a weighted sum of the log-likelihoods of the initial observations. Since the weights of the signal likelihoods coincide with the network centralities of their respective agents, these weights can be equalized in degree-regular and balanced topologies, where all nodes have the same in and out degrees. Therefore, in such highly symmetric structures the support of the consensus belief coincides with the maximum likelihood estimators (MLE) of the truth state. The latter signifies a measure of efficiency in balanced regular structures; nevertheless, the asymptotic beliefs systematically reject the less probable alternatives in spite of the limited initial data, and in contrast with the optimal (Bayesian) belief of an observer with complete information of the environment and private signals. The latter would assign probabilities proportionally to the likelihood of every state, without rejecting any of the possible alternatives. The asymptotic rejection of less probable alternatives indicates a case of group polarization, i.e. overconfidence in the group aggregate. Unlike the linear action updates and the DeGroot model which entail a host of knife-edge conditions on the signal structure and model parameters, we observe that the belief updates are unweighted; not only they effectively internalize the heterogeneity of the private observations, but also they compensate for the individual priors. Thence, we are lead to the conclusion that multiplicative belief updates, when applicable, provide a relatively robust description of the decision making behavior.

Bio: M. Amin Rahimian is a recipient of gold medal in 2004 Iran National Chemistry Olympiad. He was awarded an honorary admission to Sharif University of Technology, where he received his B.Sc. in Electrical Engineering-Control. In 2012 he received his M.A.Sc. from Concordia University in Montréal, and in 2016 he received his A.M. in Statistics from the Wharton School at the University of Pennsylvania, where he is currently a PhD student at the Department of Electrical and Systems Engineering and the GRASP Laboratory. He was a finalist in 2015 Facebook Fellowship Competition, as well as 2016 ACC Best Student Paper Competition. His research interests include network science, distributed control and decision theory, with applications to social and economic networks.

September 21
Alec Koppel
Sparse Online Task-Driven Kernel Learning
Click here to view Alec Koppel's seminar

Read the Abstract and Bio

Abstract: We consider stochastic nonparametric regression problems in a reproducing kernel Hilbert space (RKHS), an extension of the expected risk minimization framework popular in supervised machine learning to nonlinear function estimation. Conventional wisdom is that kernel methods are inapplicable to supervised learning problems where the number of training examples is possibly infinite, since the application of stochastic gradient methods which alleviates the bottleneck in the training sample size for standard expected risk minimization suffers from the fact that in the RKHS setting the decision function is parameterized by a coefficient vector and kernel matrix whose size is at least proportional to the iteration index. We propose a method which alleviates the computational intractability of kernelized expected risk minimization in two ways: (1) we consider the use of functional stochastic gradient method (FSGD) which operates on stochastic functional gradients in lieu of true gradients, and hence is able to operate on a subset of training examples at each step; and (2), we reduce the dimensionality of the FSGD sequence in a manner that guarantees the resulting functions admit parsimonious parameterizations. Dimensionality reduction of the FSGD sequence is achieved by projecting the iterates onto subspaces of the Hilbert space that are spanned by a sparse subset of past kernel evaluations of training examples (model points) that are most important for stochastic descent. To obtain these sparse subspaces, we make use of a greedy sub-routine called kernel orthogonal matching pursuit (KOMP) which searches through the instantaneous kernel dictionary and discards as many model points as possible for a fixed approximation budget. We establish the almost-sure convergence of this method, which we call Parsimonious Online Learning with Kernels (POLK), in both the diminishing and constant algorithm step-size regimes when the respective sparse approximation budget is chosen as the square of the step-size or the square-root of the cube of the step-size. These results require that the loss function is convex and Holder continuous, a particular regularization parameter selection, standard stochastic gradient variance conditions, and compactness of the feature space. We further show that the task- driven kernel dictionary which parameterizes the function sequence is guaranteed to be finite, and the model order depends on loss function smoothness properties as well as the sparse approximation budget of KOMP. The method is then numerically evaluated for a kernel multi-class support vector classification task on a synthetic Gaussian mixture model dataset and the MNIST hand-written digits. The method exhibits a state of the art trade off between complexity and accuracy on these tasks.

Bio: Alec Koppel is a doctoral student in the Department of Electrical and Systems Engineering at the University of Pennsylvania and a participant in the Science, Mathematics, and Research for Transformation (SMART) Scholarship Program sponsored by the American Society of Engineering Education. His sponsoring facility is the U.S. Army Research Laboratory in Adelphi, MD, where he works during doctoral summers. Before coming to Penn, completed his M.S. degree in Systems Science and Mathematics and B.A. degree in Mathematics at Washington University in St. Louis, MO. His research interests are in the areas of signal processing, optimization, and learning theory. His current work focuses on designing new large-scale or dynamic optimization methods for multi-agent systems with a focus on robotic and computer networks.

September 28
Xilin Liu
A Closed-loop Bidirectional Brain-Machine Interface System for Freely Behaving Animals
Click here to view Xilin Liu's seminar

Read the Abstract and Bio

Abstract: Brain machine interface (BMI) creates an artificial pathway between the brain and the external world. The researches and applications of BMI receive enormous attention among scientific community and public in the past decade. However, most existing research of BMI relies on experiments of tethered or sedated animals, using rack-mount equipment, which significantly restricts the accessible experiments. Moreover, the vast majority of research to date has focused on neural signal recording and decoding in an open-loop method. Though the use of closed-loop methods with wireless BMI is critical to the success of an extensive of neuroscience research, it is an approach yet to be widely used, with the electronics design being one of the major bottlenecks. The key goal of this research is to address the design challenges of a closed-loop, bidirectional BMI by providing innovative solutions from neuron-electronics interface up to the system level. Prototype devices are being designed, fabricated and verified in bench testing. In-Vivo experiments have been conducted using the developed BMI system. Long term recording and stimulation have been tested in animals during freely behavior, including awake, sleep (Monkeys) and swimming (Rats). Methods, circuit techniques and system topology proposed in this work can be used in a wide range of relevant neurophysiology research, especially applications of brain machine interface in freely behaving animals.

Bio: Xilin Liu (S'13) received the B.S. degree in electrical engineering from the Harbin Institute of Technology, China, in 2011, and M.S. degree in electrical engineering from the University of Pennsylvania in 2013. He is currently working toward the Ph.D. degree at the University of Pennsylvania. His research interests include analog and mixed-signal integrated circuits and systems design for medical applications, brain-machine interface, low-power data converters, high-speed DAC for telecommunications, and CMOS image sensors. He received the IEEE Solid-State Circuits Society (SSCS) 2015-16 predoctoral achievement award. He received the Best Paper Award (1st place) of the 2015 Biomedical Circuits and Systems Conference (BioCAS), and the Best Paper Award of the BioCAS Track of the 2014 International Symposium on Circuits and Systems (ISCAS). He is also the recipient of the Student-Research Preview Award (Honorable Mention) of the 2014 IEEE International Solid-State Circuits Conference (ISSCC). 

October 5
Bhoram Lee
Online Self-Supervised Learning for Robotic Visual Perception

Click here to view Bhoram Lee's seminar

Read the Abstract and Bio

Abstract: Perceiving the world has been one of the long-standing problems in computer vision and robotic intelligence. Whereas human perception works so effortlessly, even state-of-the-art algorithms experience difficulty in performing the same tasks. In this talk, I will talk about an online self-supervised learning framework for robotic visual perception, and present two case studies. In order to enable high-level or interactive tasks in the real 3D world, both semantic (‘what’) and spatial (‘where’) scene understanding are critical. Humans are known to have two distinct visual processing systems called  and  dorsal and stream, which are often called ‘what’ and ‘where’ pathways respectively. As opposed to conventional approaches in computer vision that have parallelized the two issues, my study is motivated by the interaction between the two systems.

I will present a study on monocular vision-based ground surface estimation and classification. The ground is the most important background object, which appears everywhere if on land. Being ubiquitous, the ground exhibits diverse visual features depending where you are and when it is. In this study, an online simultaneous geometric estimation and appearance-based classification of the ground is demonstrated using a large-scale dataset developed for autonomous driving car research. I will also talk about an online self-supervised approach for 3D object tracking. Knowing the precise 3D pose of an object is crucial for interactive robotic tasks such as grasping and manipulation. The three complementary modules of shape, appearance, and motion of our framework enable the self-supervision mechanism to work without any pretraining. Our approach outperformed other algorithms on public datasets, and a qualitative investigation on various scenes shows the effectiveness of our approach.

Bio: Bhoram Lee is a PhD candidate in the Department of Electrical and Systems Engineering as well as  the GRASP (General Robotics, Automation, Sensing, and Perception) Lab, at the University of Pennsylvania, under the supervision of Prof. Daniel D. Lee. Before coming to Penn, she worked at SAIT (Samsung Advanced Institute of Technology) from 2007 to 2013 as a researcher. She received B.S. in mechanical and aerospace engineering in 2005 and M.S. in aerospace engineering in 2007 from Seoul National University (SNU), Korea. Her previous research experience includes visual navigation of aerial robots, sensor fusion, and mobile user interfaces. Her current academic interest includes computer vision, machine learning, and general robotics with a focus on improving robotic perception via online learning approaches.

October 12
David Hopper
High-Fidelity Readout of a Room Temperature Solid-State Qubit
Click here to view David Hopper's Seminar

Read the Abstract and Bio

Abstract: The rapidly emerging field of quantum information processing promises advances in a diverse set of application areas. A few examples of note are the potential for exponential speed ups over classical computers, provide fundamentally secure communication channels, and create unparalleled nanoscale environmental sensors. Critical to all of these applications is the possession of a qubit that can be initialized, controlled, entangled, and readout with great precision. The nitrogen-vacancy (NV) center in diamond has become a workhorse in proof-of-principle demonstrations of many of these applications in recent years. While the NV center offers all-optical mechanisms for initialization and readout, both are imperfect. These imperfections lead to the need to repeat experiments more than 10,000 times to obtain adequate signal-to-noise ratios. Furthermore, the traditional readout protocol is unsuited to data post-processing strategies. In this talk, I will overview the current state-of-the art in NV center readout techniques, and discuss recently demonstrated protocols by our group for all-optically improving both the initialization and readout of NV centers through multicolor optical control. In addition, I will discuss how proper choice of post-processing techniques can lead to improved gains in measurement accuracy and efficiency.
Bio: David Hopper earned his bachelor’s degree in Physics with Honors from The Pennsylvania State University in 2014. His research was on developing exfoliation and fabrication techniques for creating low-dimensional topological insulator devices. He is currently pursuing his PhD in Physics at the University of Pennsylvania with Professor Lee Bassett as part of the physics graduate group program. He is broadly interested in semiconductor quantum dynamics and quantum information science, with a current focus on improving the readout and storage capabilities of the nitrogen-vacancy center in diamond.

October 19
Luiz Chamon
Greedy Sampling of Graph Signals

Read the Abstract and Bio

Abstract: Graph signal processing (GSP) is an emerging field that investigates signals supported on irregular domains. It provides an abstraction that enables the study under a unified framework of problems from subspace tracking, network processes, and statistics. It has found applications in sensor networks, image processing, clustering, and neuroscience, to name a few. Given its fundamental role in signal processing and statistics, it is not surprising that sampling has attracted considerable interest from the GSP community. In contrast to traditional signal processing, however, the irregularity of the signal domain makes the selection of the sampling points nontrivial and NP-hard in general. In practice, greedy sampling remains ubiquitous and has been successfully used in several application. Nevertheless, it is well-known that the mean-square error (MSE) is not supermodular and no alternative performance analysis has been derived to justify using this method. This work sets out to explain the success of greedy sampling by introducing the concept of approximate supermodularity and updating the classical greedy bound for this class of functions. Then, it quantifies the approximate supermodularity of typical figures of merit, including the MSE, showing that they can be optimized with worst-case guarantees using greedy search.

Bio: Luiz Chamon is a PhD candidate in the Department of Electrical and Systems Engineering at the University of Pennsylvania under the supervision of Prof. Alejandro Ribeiro. He completed his electrical engineering bachelor and masters in 2015 at the University of São Paulo, Brazil. During this time, his research involved adaptive filtering, acoustic MIMO equalization, and low complexity decimation/interpolation structures. He also participated in several projects involving acoustical design and electronics for audio signal processing and provided statistical analysis consulting in areas such as psychology and ergonomics. Currently, his research interests lie on topics in optimization and graph signal processing.


October 26

Read the Abstract and Bio


November 2
Fernando Gama
Rethinking Sketching as Sampling: A Graph Signal Processing Approach

Read the Abstract and Bio

Abstract: Sampling of bandlimited graph signals has well-documented merits for dimensionality reduction, affordable storage, and online processing of streaming network data. Most existing sampling methods are designed to minimize the error incurred when reconstructing the original signal from its samples. Oftentimes these parsimonious signals serve as inputs to computationally-intensive linear transformations (e.g., graph filters). Hence, interest shifts from reconstructing the signal itself towards instead approximating the output of the prescribed linear operator efficiently. In this context, we propose a novel sampling scheme that leverages the bandlimitedness of the input as well as the transformation whose output we wish to approximate. We formulate problems to jointly optimize sample selection and a sketch of the target linear transformation, so when the latter is affordably applied to the sampled input signal the result is close to the desired output. The developed sampling plus reduced-complexity processing pipeline is particularly useful for streaming data, where the linear transform has to be applied fast and repeatedly to successive inputs.

Bio: Fernando Gama is currently a doctoral student in the Department of Electrical and Systems Engineering at the University of Pennsylvania. His advisor is Alejandro Ribeiro. He has received a Fulbright Scholarship for International Students. He obtained his Electronic Engineering Degree at the University of Buenos Aires, Argentina in 2013. His research interests are in the Signal Processing field and its application to Network Theory.

November 9
Santiago Paternain
Navigation Functions for Convex Potentials in Spaces with Convex Obstacles

Read the Abstract and Bio

Abstract: In the pursue of autonomous behavior we need to endow robots with the ability to perform complex tasks in complex environments. In particular, we advocate to the resolution of problems where the goal configuration is not given explicitly to the agent but provided as a high level task defined by the means of an objective function that must be minimized or maximized. Several problems in robotics such as hill climbing and source seeking fall into this category and while solutions to these particular problems have been provided a unifying framework and theoretical guarantees are still to be developed in cases where the environment presents obstacles to the robot. We consider an artificial potential of the Rimon-Koditschek form and we provide theoretical guarantees on the geometry of the space for which such potential is a navigation function. Furthermore, we provide theoretical guarantees that a stochastic version of such approach succeeds in driving the robot towards the solution under the same geometrical conditions than in the deterministic scenario.

Bio: Santiago Paternain received the B.Sc. degree in electrical engineering from Universidad de la República Oriental del Uruguay, Montevideo, Uruguay in 2012. Since August 2013, he has been working toward the Ph.D. degree in the Department of Electrical and Systems Engineering, University of Pennsylvania. His research interests include optimization and control of dynamical systems.

November 16
Zhe Xuan
High-Speed Optical Receiver in Silicon for Next-Generation Networks

Read the Abstract and Bio

Abstract: Optical interconnects continue to play a key role for high capacity data transfer in data-centers, supercomputers, and more recently as ultra-high data-rate chip-to-chip interconnects. The rapidly growing capacity demand has pushed the serial data-rate of optical links to 40 Gb/s and beyond. The compatibility with state of the art high data-rate systems such as microprocessors and memory banks in conjunction with recent advances in silicon photonics have made low-cost Si-based transceivers an attractive candidate for massive deployment. In this talk, a low-power 40 Gb/s optical receiver is presented where silicon photonic techniques and careful electromagnetic simulations are used to design and fabricate a SiGe waveguide-coupled p-i-n photodiode (PD) in a 0.18 μm Ge-on-SOI process and an electronic circuitry in a 0.13 μm SiGe BiCMOS process. The co-design of the PD, the electronic circuitry have enabled realization of a compact mm-wave optical receiver with low-cost and repeatable packaging solution using bond-wires. The integrated optical receiver works at 40 Gb/s for 231-1 pseudorandom binary sequence (PRBS) while consuming 77 mW.

Bio: Zhe Xuan received the B.S. degree in Electrical Engineering from Nanjing University, China, in 2011. He is currently working towards the Ph.D. in Electrical Engineering under the supervising of Prof. Firooz Aflatouni. His research interests include on-chip optical devices, energy efficient transceiver front-ends, and novel optical systems.


November 23

Read the Abstract and Bio


November 30
Markos Epitropou
Competitive Equilibrium and Trading Networks: A Network Flow Approach

Read the Abstract and Bio

Abstract: Under full substitutability of preferences, it has been shown that a competitive equilibrium exists in trading networks, and is equivalent (after a restriction to equilibrium trades) to (chain) stable outcomes. In this paper, we formulate the problem of finding an efficient outcome as a generalized submodular flow problem on a suitable network. Equivalence with seemingly weaker notions of stability follows directly from the optimality conditions, in particular the absence of improvement cycles in the flow problem. Our formulation yields strongly polynomial algorithms for finding competitive equilibria in trading networks, and testing (chain) stability.

Bio: Markos Epitropou received his Diploma in Electrical and Computer Engineering from the National Technical University of Athens, Athens, Greece in 2012. He later received a Ms. S. in Logic, Algorithms and Computation from the National and Kapodistrian University of Athens, Athens, Greece in 2014. Since August 2014, he has been working towards the Ph.D. degree in the Department of Electrical and Systems Engineering, University of Pennsylvania. His research interests include Market Design, Algorithm Design, and Combinatorial Optimization.


December 7

Read the Abstract and Bio


December 14
Timothy Jones
Quantitative microwave impedance microscopy with effective medium approximations

Read the Abstract and Bio

Abstract: Microwave impedance microscopy (MIM) is a mode of atomic force microscopy (AFM) that measures changes in electrical admittance with nanoscopic resolution. The high resolution of this local measurement has made it possible to investigate several systems currently of interest in applied physics: the highly conducting domain walls in certain ferromagnetic materials, quantum hall edge states in two-dimensional electron gasses, and phase separation in strained manganites, for example. This technique can also be used to investigate local structural variation in inhomogeneous materials. While material structure is often characterized macroscopically with scattered electromagnetic fields (e.g. X-ray diffraction), the interaction between a sample and the short-range evanescent wave radiating from the MIM tip carries information on local structure. In this talk I will describe various applications of the technique and discuss efforts to quantify the MIM output and infer structural features of materials, specifically in tunable carbon structures often used in electrochemical charge storage devices.

Bio: In 2010, Tim received the B.S.E. in Electrical Engineering from the University of Pennsylvania. He has remained in the same department at Penn to study for the Ph.D. His current research involves the computational modeling of tip-sample interaction in functional atomic force microscopy techniques.