| Start Date: 04 Apr 2022; Start Time: 10:30AM - 12:00PM Title: Scalable and Provably Robust Algorithms for Machine Learning Bio: Yu Cheng is an assistant professor in the Mathematics department at the University of Illinois at Chicago. He obtained his Ph.D. in Computer Science from the University of Southern California. Before joining UIC, he was a postdoc at Duke University and a visiting member at the Institute for Advanced Study. His main research interests include machine learning, optimization, and game theory. Speaker: Abstract: As machine learning plays a more prominent role in our society, we need learning algorithms that are reliable and robust. It is important to understand whether existing algorithms are robust against adversarial attacks and design new robust solutions that work under weaker assumptions. The long-term goal is to bridge the gap between the growing need for robust algorithms and the lack of systematic understanding of robustness. In this talk, I will discuss the challenges that arise in the design and analysis of robust algorithms for machine learning. I will focus on three lines of my recent work: (1) designing faster and simpler algorithms for high-dimensional robust statistics where a small fraction of the input data is arbitrarily corrupted, (2) analyzing the optimization landscape of non-convex approaches for low-rank matrix problems and making non-convex optimization robust against semi-random adversaries, and (3) considering learning in the presence of strategic behavior where the goal is to design good algorithms that account for the agents' strategic responses. Location: Via Zoom Committee: |
| Start Date: 05 Apr 2022; Start Time: 09:00AM - 10:00AM Title: Generative Neurosymbolic Machines Bio: Speaker: Abstract: Reconciling symbolic and distributed representations is a crucial challenge that can potentially resolve the limitations of current deep learning. Remarkable advances in this direction have been achieved recently via generative object-centric representation models. While learning a recognition model that infers object-centric symbolic representations like bounding boxes from raw images in an unsupervised way, no such model can provide another important ability of a generative model, i.e., generating (sampling) according to the structure of learned world density. In this work, we propose Generative Neurosymbolic Machines, a generative model that combines the benefits of distributed and symbolic representations to support both structured representations of symbolic components and density-based generation. These two crucial properties are achieved by a two-layer latent hierarchy with the global distributed latent for flexible density modeling and the structured symbolic latent map. To increase the model flexibility in this hierarchical structure, we also propose the StructDRAW prior. In experiments, we show that the proposed model significantly outperforms the previous structured representation models as well as the state-of-the-art non-structured generative models in terms of both structure accuracy and image generation quality. Location: Via Zoom Committee: Prof. Sungjin Ahn (advisor) Prof. Hao Wang Prof. Dimitris Metaxas Prof. Sudarsun Kannan |
| Start Date: 12 Apr 2022; Start Time: 09:00AM - 10:00AM Title: Factoring and Learning Algorithms for Low-depth Algebraic Circuits. Bio: Speaker: Abstract: Polynomial factoring and learning arithmetic circuits (a.k.a. circuit-reconstruction) are two fundamental problems in algebraic complexity. They are deeply connected to the questions of circuit lower bound and polynomial identity testing (PIT). In this dissertation, we present various new results on learning and factoring of low-depth circuits, while emphasizing these connections. Some of our main results are as follows:1. Sparse Polynomial Factoring: We study the problem of deterministic factorization of sparse polynomials. We show that if $f \in \Fn$ is a polynomial with $s$ monomials, with individual degrees of its variables bounded by $d$, then $f$ can be deterministically factored in time $s^{\poly(d)\log n}$. A crucial ingredient in our proof is a quasi-polynomial sparsity bound for factors of sparse polynomials of bounded individual degree. In particular we show if $f$ is an $s$-sparse polynomial in $n$ variables, with individual degrees of its variables bounded by $d$, then the sparsity of each factor of $f$ is bounded by $s^{\BigO({d^2\log{n}})}$. 2. Learning low-rank tensors and depth 3 multilinear circuits: We give new and efficient black-box reconstruction algorithms for some classes of depth-$3$ arithmetic circuits. As a consequence, we obtain the first efficient algorithm for computing the tensor rank and for finding the optimal tensor decomposition as a sum of rank-one tensors when the input is a {\em constant-rank} tensor.3. Learning depth-4 multilinear circuits over finite Fields: We present a deterministic algorithm for reconstructing multilinear $\Spsp(k)$ circuits, i.e. multilinear depth-$4$ circuits with fan-in $k$ at the top $+$ gate. For any fixed $k$, given black-box access to a polynomial $f \in \Fn$ computable by a multilinear $\Spsp(k)$ circuit of size $s$, the algorithm runs in time quasi-poly($n,s,\size{\F})$ and outputs a multilinear $\Spsp(k)$ circuit of size quasi-poly($n,s$) that computes $f$.4. Average-case learning for generalized depth-3 circuits: We design a (randomized) learning algorithm for \textit{generalized} depth-3 circuits. A circuit in this class is an expression of the type, $g_1(\ell_{11}, \ldots, \ell_{1m}) + \cdots + g_s(\ell_{s1}, \ldots, \ell_{sm}),$ where $g_i$'s are $m$-variate degree $d$ homogeneous polynomials and $\ell_{ij}$'s are linear forms in the variables $\vecx = (x_1,\ldots, x_n)$. Our work is in line with the ideas presented in \cite{KayalSaha19, GargKayalSaha20} where they use lower bound methods in arithmetic complexity to design average-case learning algorithms. Location: Via Zoom Committee: Shubhangi Saraf (chair) Eric Allender Sepehr Assadi Zeev Dvir (external) |
| Start Date: 22 Apr 2022; Start Time: 04:00PM - 05:00PM Title: Instance Dependent Sample Complexity Bounds for Interactive Learning Bio: Kevin Jamieson is an Assistant Professor in the Paul G. Allen School of Computer Science & Engineering at the University of Washington and is the Guestrin Endowed Professor in Artificial Intelligence and Machine Learning. He received his B.S. in 2009 from the University of Washington, his M.S. in 2010 from Columbia University, and his Ph.D. in 2015 from the University of Wisconsin - Madison under the advisement of Robert Nowak, all in electrical engineering. He returned to the University of Washington as faculty in 2017 after a postdoc with Benjamin Recht at the University of California, Berkeley. Jamieson's work has been recognized by an NSF CAREER award and Amazon Faculty Research award. His research explores how to leverage already-collected data to inform what future measurements to make next, in a closed loop. The work ranges from theory to practical algorithms with guarantees to open-source machine learning systems and has been adopted in a range of applications, including measuring human perception in psychology studies, adaptive A/B/n testing in dynamic web-environments, numerical optimization, and efficient tuning of hyperparameters for deep neural networks. Speaker: Abstract: The sample complexity of an interactive learning problem, such as multi-armed bandits or reinforcement learning, is the number of interactions with nature required to output an answer (e.g., a recommended arm or policy) that is approximately close to optimal with high probability. While minimax guarantees can be useful rules of thumb to gauge the difficulty of a problem class, algorithms optimized for this worst-case metric often fail to adapt to “easy” instances where fewer samples suffice. In this talk, I will highlight some my group’s work on algorithms that obtain optimal, finite time, instance dependent sample complexities that scale with the true difficulty of the particular instance, versus just the worst-case. In particular, I will describe a unifying experimental design based approach used to obtain such algorithms for best-arm identification for linear bandits, contextual bandits with arbitrary policy classes, and smooth losses for linear dynamical systems. Location: Via Zoom Committee: |
| Start Date: 28 Apr 2022; Start Time: 04:00PM - Title: The Human in Human-Robot Interaction Bio: Dr. Matthew Gombolay is an Assistant Professor of Interactive Computing at the Georgia Institute of Technology. He received a B.S. in Mechanical Engineering from the Johns Hopkins University in 2011, an S.M. in Aeronautics and Astronautics from MIT in 2013, and a Ph.D. in Autonomous Systems from MIT in 2017. Gombolay’s research interests span robotics, AI/ML, human-robot interaction, and operations research. Between defending his dissertation and joining the faculty at Georgia Tech, Dr. Gombolay served as technical staff at MIT Lincoln Laboratory, transitioning his research to the U.S. Navy and earning an R&D 100 Award. His publication record includes a best paper award from the ACM/IEEE International Conference on Human-Robot Interaction, and the American Institute for Aeronautics and Astronautics, a finalist for best paper at the Conference on Robot Learning, and a finalist for best student paper at the American Controls Conference. Dr Gombolay was selected as a DARPA Riser in 2018, received 1st place for the Early Career Award from the National Fire Control Symposium, and was awarded a NASA Early Career Fellowship. This talk is organized by the Rutgers SOCRATES project. Speaker: Abstract: New advances in robotics and autonomy offer a promise of revitalizing final assembly manufacturing, assisting in personalized at-home healthcare, and even scaling the power of earth-bound scientists for robotic space exploration. Yet, in real-world applications, autonomy is often run in the O-F-F mode because researchers fail to understand the human in human-in-the-loop systems. In this talk, I will share exciting research we are conducting at the nexus of human factors engineering and cognitive robotics to inform the design of human-autonomy interaction. In my talk, I will focus on our recent work (1) democratizing robot learning by formulating better models of heterogeneous and suboptimal human teachers; (2) explaining to those humans what the robot has learned to facilitate shared mental models; and (3) scaling the human-robot interaction to team-level coordination with graph neural networks. The goal of this research is to inform the design of autonomous teammates so that users want to turn – and benefit from turning – autonomy to the O-N mode. Location: Via Zoom Committee: |
| Start Date: 29 Apr 2022; Start Time: 04:00PM - 05:00PM Title: Explaining the Decisions of AI Systems Bio: Adnan Darwiche is a professor and former chairman of the computer science department at UCLA. He directs the Automated Reasoning Group, which focuses on symbolic reasoning, probabilistic reasoning and their applications to machine learning. Professor Darwiche is Fellow of AAAI and ACM and recipient of the Lockheed Martin Excellence in Teaching Award. He is a former editor-in-chief of the Journal of Artificial Intelligence Research (JAIR) and author of "Modeling and Reasoning with Bayesian Networks," by Cambridge University Press. Presented in association with the DATA-INSPIRE TRIPODS Institute. Speaker: Abstract: I will present a theory for reasoning about the decisions made by AI systems, particularly classifiers such as decision trees, random forests, Bayesian networks and some limited types of neural networks. The theory is based on “compiling" the input-output behavior of classifiers into discrete functions in the form of tractable circuits. At the heart of the theory is the notion of “complete reason” behind a decision which is extracted from a circuit-instance pair and can be used to answer many queries about the decision, including ones pertaining to explainability, robustness and bias. I will also overview developments on tractable circuits which provide the computational arm for employing this theory in practice and will briefly overview recent results on quantified Boolean logic which provide classifier-independent semantics of this theory that further broadens its applicability. Location: Presented via Zoom Committee: |
| Start Date: 02 May 2022; Start Time: 02:00PM - 04:00PM Title: Beyond Instance-level Reasoning in Object Pose Estimation and Tracking for Robotic Manipulation Bio: Speaker: Abstract: This thesis deals with object pose estimation and tracking, and solve robot manipulation tasks. It aims to address uncertainty due to dynamics and generalize to novel object instances by reducing the dependency on either instance or category level 3D models. Robot object manipulation often requires reasoning about object poses given visual data. For instance, pose estimation can be used to initiate pick-and-drop manipulation and has been studied extensively. Purposeful manipulation, however, such as precise assembly or withinhand re-orientation, requires sustained reasoning of an object's state, since dynamic effects due to contacts and slippage, may alter the relative configuration between the object and the robotic hand. This motivates the temporal tracking of object poses over image sequences, which reduces computational latency, while maintaining or even enhancing pose quality relative to single-shot pose estimation. Most existing techniques in this domain assume instance-level 3D models. This complicates generalization to novel, unseen instances, and thus hinders deployment to novel environments. Even if instance-level 3D models are unavailable, however, it may be possible to access category-level models. Thus, it is desirable to learn category-level priors, which can be used for the visual understanding of novel, unknown object instances. In the most general case, where the robot has to deal with out-of-distribution instances or it cannot access category-level priors, object-agnostic perception methods are needed. Given this context, this thesis proposes a category-level representation, called NUNOCS, to unify the representation of various intra-class object instances and facilitate the transfer of category-level knowledge across such instances. This work also integrates the strengths of both modern deep learning as well as pose graph optimization to achieve generalizable object tracking in the SE(3) space, without needing either instance or category level 3D models. When instance-level object models are available, a synthetic data generation pipeline is developed to learn the relative motion along manifolds by reasoning over image residuals. This allows to achieve state-of-art SE(3) pose tracking results, while circumventing manual efforts in data collection or annotation. It also demonstrates that the developed solutions for object tracking provide efficient solutions to multiple manipulation challenges. Specifically, this thesis starts from a single-image object pose estimation approach that deals with severe occlusions during manipulation. It then moves to long-term object pose tracking via reasoning over image residuals between consecutive frames, while training exclusively over synthetic data. In the case of object tracking along a video sequence, the dependency on either instance-level or category-level CAD models is reduced via leveraging multi-view consistency, in the form of a memory-augmented pose graph optimization, to achieve spatial-temporal consistency. For initializing pose estimates in video sequences involving novel unseen objects, category-level priors are extracted by taking advantage of easily accessible virtual 3D model databases. Following these ideas, frameworks for category-level, task-relevant grasping, and vision-based, closed-loop manipulation are developed, which resolve complicated and high precision tasks. The learning process is scalable as the training is performed exclusively over synthetic data or through a robot's self-interaction process conducted solely in simulation. The proposed methods are evaluated first over public computer vision benchmarks, boosting the previous state-of-art tracking accuracy from 33.3% to 87.4% on the NOCS dataset, despite reducing dependency on category-level 3D models for training. When applied to real robotic setups, they significantly improve category-level manipulation performance, validating their effectiveness and robustness. In addition, this thesis unlocks and demonstrates multiple complex manipulation skills in open world environments. This is despite limited input assumptions, such as training solely over synthetic data, dealing with novel unknown objects, or learning from a single visual demonstration. Location: 1 Spring St, New Brunswick, NJ 08901 Rm 319 Committee: Kostas Bekris (advisor) Abdeslam Boularias Dimitris Metaxas Shuran Song (Columbia University - external) |
| Start Date: 05 May 2022; Start Time: 10:00AM - 11:30AM Title: Characterizations of Experts Algorithms That Are Incentive Compatible Bio: Speaker: Abstract: Experts algorithms are online learning algorithms that combine the predictions of multiple experts to produce an aggregate prediction. These algorithms have strong accuracy guarantees: the loss of the aggregate is at most a small amount more than the loss of the best expert. Most analyses assume that the experts volunteer their best information completely and truthfully. Experts are treated like thermometers reporting temperatures, without any agency of their own. Recently, authors have explored new algorithms that take the experts’ motivations into account. They treat experts as rational, not altruistic: experts maximize their own influence on the algorithm and thus their own reputations as forecasters. In this paper, we look at many standard expert algorithms and ask what happens if experts act strategically. We prove several sufficient conditions under which standard expert algorithms are incentive compatible, meaning that the expert’s best strategy is to report their true belief about the realization of each event. Some conditions lead to incentive compatibility in the limit of the number of experts, while other more specific conditions produce incentive compatibility regardless of the number of experts and across all external parameters on the expert algorithms. We illustrate our main results through simulation. Location: Via Zoom Committee: David Pennock (advisor) Sepehr Assadi Yongfeng Zhang Desheng Zhang |
| Start Date: 09 May 2022; Start Time: 04:00PM - Title: Improved Bounds for Distributed Load Balancing Bio: Speaker: Abstract: We consider the problem of load balancing in the distributed setting. The input is a bipartite graph of clients and servers where each client comes with a positive weight. The goal is to assign each client to one of its adjacent servers minimizing the weighted sum of clients assigned to any one server. This problem has a variety of applications and has been widely studied under several different names, including scheduling with restricted assignment, semi-matching, and distributed backup placement. We give the first LOCAL algorithm to compute an O(1)-approximation to the load balancing problem in polylog(n) rounds. In the CONGEST model, we give an O(1)-approximation algorithm in polylog(n) rounds when the client weights are uniform; for general weights, the approximation ratio is O(log(n)). Based on joint work with Sepehr Assadi and Aaron Bernstein. Location: Via Zoom Committee: Sepehr Assadi (Advisor) Aaron Bernstein Martin Farach-Colton Vladimir Pavlovic |
| Start Date: 12 May 2022; Start Time: 02:30PM - 04:30PM Title: Interleaving Learning and Search for Solving Long-Horizon Episodic Robotic Planning Tasks Bio: Speaker: Abstract: Robots need to compute high-quality solutions or plans quickly to provide meaningful assistance in solving household and industrial tasks. However, computing high-quality, long-horizon plans are rather challenging. In my research, I attempt to harness the power of deep learning, reinforcement learning, and Monte Carlo tree search to tackle long-horizon robot planning tasks. This presentation will discuss our proposed solutions, employing the tools mentioned earlier that effectively solve two long-horizon robot manipulation problems: de-clutter and object retrieval. Location: Virtual Committee: Proffesor Jinjin Yu (Advisor) Professor Abdeslam Boularias Professor Kostas Bekris Professor Karthik Srikanta
|
| Start Date: 18 May 2022; Start Time: 10:00AM - Title: Machine Learning and Understanding Art Bio: Speaker: Abstract: Recent advances in machine learning on various computer vision tasks have shown the great potential for developing an AI system for art through the successful applications in prediction of style, genre, medium, attribution, school of art, etc. Beyond the categorical information, in this dissertation, a more fundamental level of artistic knowledge is pursued by developing machine learning systems for art principles. The art principles are to know how art is visually formed and identify what content in it and what symbolic meaning it has. The task necessitates fine-grained semantics to describe art, but art data does not generally accommodate the fine details. The scarcity of data annotation is a primary challenge in this machine learning study. Three research problems are explored; (1) first is to find principal semantics for style recognition. (2) second is to lay the groundwork for computational iconography, i.e. recognize content and discover the co-occurrence and visual similarities among the content in fine art paintings. (3) third is to quantify paintings with finite visual semantics from style through language models. In the system design, well-established knowledge and facts in art theory are leveraged, or general knowledge of art is integrated into the hierarchical architecture of deep-CNN as a numerical form after learning it from a corpus of art-texts through contemporary language models in Natural Language Processing. Language modeling is a practical and scalable solution requiring no direct annotation, but it is inevitably imperfect. This dissertation shows how deep learning's hierarchical structure and adaptive nature can create a stronger resilience to the incompleteness of the practical solution than other related methods. Location: Virtual Committee: Professor Ahmed Elgammal (Advisor) Professor Konstantinos Michmizos Professor Karl Stratos Professor Gennaro Vessio (University of Bari, Italy) |
| Start Date: 24 May 2022; Start Time: 10:00AM - 11:30AM Title: Towards Fairer Recommender Systems through Deep Reinforcement Learning Bio: Speaker: Abstract: The issue of fairness in recommendation is becoming increasingly essential as Recommender Systems (RS) touch and influence more and more people in their daily lives. While most of the existing fairness-aware recommendation approaches do alleviate the impact of unfair recommendations, they also have two major drawbacks. First, most of them have been situated in a static or one-shot setting, where the protected groups of items are fixed, and the model provides a one-time fairness solution based on fairness-constrained optimization, which fails to consider the dynamic nature of the recommender systems, where attributes such as item popularity may change over time due to the recommendation policy and user engagement. Second, they mainly aim at solving a constrained optimization problem by imposing a constraint on the level of fairness while optimizing the main recommendation objective, which may significantly compromise the recommendation accuracy due to the inherent trade-off between fairness and utility. Regarding to these questions, we propose two fairness-aware recommendation frameworks leveraging the advantages of deep reinforcement learning to overcome the above drawbacks correspondingly. For the first, we tackle it by proposing a fairness-constrained reinforcement learning algorithm for recommendation, which models the recommendation problem as a Constrained Markov Decision Process (CMDP), so that the model can dynamically adjust its recommendation policy to make sure the fairness requirement is always satisfied when the environment changes. For the second, we propose a fairness-aware recommendation framework using multi-objective reinforcement learning (MORL), called MoFIR (pronounced “more fair”), which is able to learn a single parametric representation for optimal recommendation policies over the space of all possible preferences between fairness and utility. Location: Virtual Committee: Professor Yongfeng Zhang (Advisor) Professor Desheng Zhang Professor Jie Gao Professor He Zhu |
| Start Date: 25 May 2022; Start Time: 10:00AM - 11:30PM Title: Counterfactual Explainable AI Bio: Speaker: Abstract: By providing explanations for users and AI system designers to facilitate better decision making and system understanding, explainable AI has been an important research problem. It is challenging because the most state-of-the-art machine learning models, which utilize deep neural networks, are non-transparent. In my research on Counterfactual Explainable AI, we take insights of counterfactual reasoning from causal inference to address this challenge. We first mathematically formulate the complexity and strength of explanations, and then propose a general model-agnostic counterfactual learning framework to seek simple (low complexity) and effective (high strength) explanations for the model decision. More specifically, counterfactual reasoning asks, “if A did not happen, will B happen?”. When applied in the machine learning field, it looks for a minimal change on the input such that the prediction will be different. Therefore, the changed factors are crucial for the original prediction made by the system, which constitute the counterfactual explanations. We formulate this idea as a machine learning optimization problem and generate faithful explanations. Meanwhile, we also design metrics based on counterfactual reasoning to quantitatively evaluate the necessity and sufficiency of the explanations. They are proved to be suitable for evaluating generated explanations in explainable AI, which is a challenging task in this field. We conduct a comprehensive set of experiments for different machine learning applications, such as recommendations and graph-based drug mutagenicity predictions, to show the effectiveness of the proposed counterfactual-based explainable models as well as the evaluation metrics. Location: Virtual Committee: Professor Yongfeng Zhang (Advisor) Professor Desheng Zhang Professor Dong Deng Professor Jie Gao |
| Start Date: 25 May 2022; Start Time: 01:00PM - 03:00PM Title: Brain-informed Deep Learning of Human Movements with Neurophysiological Interpretations Bio: Speaker: Abstract: The accurate and reliable decoding of movement from non-invasive electroencephalography (EEG) is essential for informing therapeutic interventions ranging from neurorehabilitation robots to neural prosthetics. However, the main caveats of EEG, namely its low spatial resolution and ill-defined source localization, hinder its reliable decoding, despite progress in both statistical and the recent machine learning methods. In this talk, I will present our brain-informed deep learning solutions for the accurate and reliable decoding of movements from EEG with neurophysiological interpretations. First, I will present a 3-dimensional convolutional neural network (3D-CNN) that, when designed and trained upon the brain's constraints, can decode EEG to predict the movement primitives, namely the reaction time (RT), movement intent, and the direction of the movement. When validated on data acquired from in-house IRB-approved motor experiments, our proposed method outperforms the state-of-the-art deep networks by up to 6.74%. Next, I will show how the deep learning framework can be extended to assess the cognitive engagement (CE) of the subjects when performing a rehabilitation task. Specifically, our method uses i) a deep network that predicts the level of CE for two classes- cognitively engaged vs. disengaged with 88.13% accuracy; and ii) a novel sliding window method that predicts continuous levels of CE in real-time. Next, to widen the application domain of our method to portable brain-computer interfaces, I will present an energy-efficient neuromorphic solution to EEG decoding. Our neuromorphic solution, where a trained spiking neural network (SNN) is deployed on Intel's Loihi neuromorphic chip, achieves the same level of performance as the deep neural networks (DNN) while consuming 95% less energy per inference. Lastly, to guarantee the reliability of our solutions in real-world applications, I will present interpretation techniques that establish correspondence between the features learned by the artificial networks and the underlying neurophysiology. Overall, our approach demonstrates the importance of biological relevance in neural networks for accurate and reliable decoding of EEG, suggesting that the real-time classification of other complex brain activities may now be within our reach. Location: Via Zoom Committee: Dr. Konstantinos Michmizos (Chair) Dr. Dimitris N. Metaxas Dr. Vladimir Pavlovic Dr. Georgios D. Mitsis (McGill University) |
| Start Date: 26 May 2022; Start Time: 03:00PM - Title: Biologically Inspired Spiking Neural Networks for Energy-Efficient Robot Learning and Control Bio: Speaker: Abstract: : Energy-efficient learning and control are becoming increasingly crucial for robots that solve complex real-world tasks with limited onboard resources. Although deep neural networks (DNN) have been successfully applied to robotics, their high energy consumption limits their use in low-power edge applications. Biologically inspired spiking neural networks (SNN), facilitated by the advances in neuromorphic processors, have started to deliver energy-efficient, massively parallel, and low-latency solutions to robotics. In this defense, I will present our energy-efficient neuromorphic solutions to robot navigation, control, and learning, using SNNs on Intel's Loihi neuromorphic processor. First, I will present a biologically constrained SNN, mimicking the brain's spatial system, solving the unidimensional SLAM problem while only consuming 1% of energy compared with the conventional filter-based approach. In addition, when extending the model to 2D environments by adding biologically realistic hippocampal neurons, the SNN formed cognitive maps in real-time and helped study the neuronal interconnectivity and cognitive functions. Next, I will show how the neuromorphic approach can be extended to high-level cognitive functions such as learning control policies. Specifically, I will present a reinforcement co-learning framework that jointly trains a spiking actor network (SAN) with a deep critic network using backpropagation to learn optimal policies for both mapless navigation and high-dimensional continuous control. Compared with state-of-the-art DNN approaches, our method results in up to 140 times less energy consumption during inference, while generating a superior successful rate on mapless navigation, and achieves the same level of performance on high-dimensional continuous control when using the population-coded spiking actor network (PopSAN). Lastly, I will present how these energy gains can further be extended to training through the development of a biologically plausible gradient-based learning framework on the neuromorphic processor. The learning method is functionally equivalent to the spatiotemporal backpropagation but solely relies on spike-based communication, local information processing, and rapid online computation, which are the main neuromorphic principles that mimic the brain. Overall, our work pushes the frontiers of SNN applications to energy-efficient robotic control and learning, and hence paves the way toward the introduction of a biologically inspired alternative solution for autonomous robots running on energy-efficient neuromorphic processors. Location: Via Zoom Committee: Konstantinos Michmizos (Advisor) Vladimir Pavlovic, Abdeslam Boularias, James Bradley Aimone (Sandia National Lab) |
| Start Date: 06 Jun 2022; Start Time: 03:00PM - 04:30PM Title: Online List Labeling: Breaking the log2n Barrier Bio: Speaker: Abstract: The list labeling problem is a classical combinatorial problem with many algorithmic applications. There has long existed a gap between the lower and upper bounds in the most algorithmically interesting part of the problem's parameter space. We present our recent results, which narrow this gap for the first time in nearly 4 decades. Location: Virtual Committee: Professor Martin Farach-Colton Professor Sepehr Assadi Professor Aaron Bernstein Professor Qiong Zhang |
| Start Date: 10 Jun 2022; Start Time: 11:00AM - Title: Near-Optimal Scalable Multi-Robot Path Planning : Algorithms, and Applications Bio: Speaker: Abstract: In Multi-Robot Path Planning (MRPP), given a graph and n robots with assigned starts and goals, we need to find a collision-free path for each robot connecting the start vertex and the goal vertex. It has many applications in video games, warehouse automation, formation control, and swarm robotics. Solving MRPP optimally in terms of makespan or sum-of-costs is NP-hard. Therefore, research on near-optimal solvers that have good scalability and acceptable suboptimality is an attractive topic. In this talk, we first propose a polynomial time algorithm with asymptotic expected sub-1.5 makespan optimality guarantee, which utilizes a novel Rubik Table result. The method scales up to 30,000+ robots on 300x300 grids. In addition, we extend the algorithm to three dimensional grids and apply it on large-scale drone swarms in both simulation and real drone experiments. Then we introduce two divide-and-conquer based heuristics that enhance the scalability of classical MRPP solvers while keeping 1.x optimality. Location: Virtual Committee: Professor Jingjin Yu Professor Kostas Bekris Professor Abdeslam Boularias Professor Peng Zhang |
| Start Date: 14 Jun 2022; Start Time: 11:00AM - Title: Robot Imagination: Affordance-Based Reasoning about Unknown Objects Bio: Gregory S. Chirikjian received undergraduate degrees from Johns Hopkins University in 1988, and a Ph.D. degree from the California Institute of Technology, Pasadena, in 1992. From 1992 until 2021, he served on the faculty of the Department of Mechanical Engineering at Johns Hopkins University, attaining the rank of full professor in 2001. Additionally, from 2004-2007, he served as department chair. Starting in January 2019, he moved to the National University of Singapore, where he is serving as Head of the Mechanical Engineering Department, where he has hired 14 new professors so far. Chirikjian’s research interests include robotics, applications of group theory in a variety of engineering disciplines, applied mathematics, and the mechanics of biological macromolecules. He is a 1993 National Science Foundation Young Investigator, a 1994 Presidential Faculty Fellow, and a 1996 recipient of the ASME Pi Tau Sigma Gold Medal. In 2008, Chirikjian became a fellow of the ASME, and in 2010, he became a fellow of the IEEE. From 2014-15, he served as a program director for the US National Robotics Initiative, which included responsibilities in the Robust Intelligence cluster in the Information and Intelligent Systems Division of CISE at NSF. Chirikjian is the author of more than 250 journal and conference papers and the primary author of three books, including Engineering Applications of Noncommutative Harmonic Analysis (2001) and Stochastic Models, Information Theory, and Lie Groups, Vols. 1+2. (2009, 2011). In 2016, an expanded edition of his 2001 book was published as a Dover book under a new title, Harmonic Analysis for Engineers and Applied Scientists. Speaker: Abstract: Today’s robots are very brittle in their intelligence. This follows from a legacy of industrial robotics where robots pick and place known parts repetitively. For humanoid robots to function as servants in the home and in hospitals they will need to demonstrate higher intelligence, and must be able to function in ways that go beyond the stiff prescribed programming of their industrial counterparts. A new approach to service robotics is discussed here. The affordances of common objects such as chairs, cups, etc., are defined in advance. When a new object is encountered, it is scanned and a virtual version is put into a simulation wherein the robot ``imagines’’ how the object can be used. In this way, robots can reason about objects that they have not encountered before, and for which they have no training using. Videos of physical demonstrations will illustrate this paradigm, which the presenter has developed with his students Hongtao Wu, Meng Xin, Sipu Ruan, and others. Location: 1 Spring Street, Room 403 Committee: |
| Start Date: 23 Jun 2022; Start Time: 12:00PM - Title: Semantic Modeling, Integration and Episodic Organization of Personal Digital Traces Bio: Speaker: Abstract: Memory plays a fundamental role in life and is critical to our everyday functioning. We use memories to maintain our personal identity, to support our relationships, to learn, and to solve problems. Today, a vast amount of tools supports digital capture of different aspects of people's lives. These tools produce a multitude of data objects, which we call Personal Digital Traces - PDTs, which can be used to help reconstruct people’s episodic memories and connect to their past personal events. This reconstruction may have several applications, from helping the recall of patients with neurodegenerative diseases to helping people remember past events and better manage their data and time if this information is used in activity-centric applications, like personal assistants.This dissertation takes steps towards supporting autobiographical memory by associating heterogeneous PDTs according to their higher-level purposes and usages and summarizing them into episodic narratives. We start by presenting a unified and intuitive conceptual modeling language whose novel features include the properties "who, what, when, where, why, how" applied uniformly to both personal digital traces and their corresponding atomic events/tasks that produce them. We then proceed by describing an ontology for prototypical higher level plans ("scripts") for common everyday events and show how families of related scripts can be defined by incrementally modifying more general scripts through inheritance. We then present a multiplayer web-based game called OneOfUs, which creates possible associated digital trace descriptions that can be produced by each of these activities/scripts through crowdsourcing techniques that act as evidence for the execution of such scripts. The game is able to automatically validate and assess knowledge at the time of the game, as well as dynamically learn new pieces of information, as it has a way of not neglecting uncommon answers through players' votes. For instantiating those scripts based on the lower level actions of which scripts are composed of, we present a bottom up merging algorithm that groups and relates several digital traces from many different sources into script instances (episodes) as well as a software architecture that supports systematic and declarative specification of evidence. This also utilizes a scoring scheme to account for the varied strength of evidence provided by PDTs or script steps.Finally, to evaluate the efficacy of our methodology, we designed and implemented YourDigitalSelf, an Android mobile device application that gathers and integrates personal digital traces into narratives. A thorough evaluation performed over real user's data collections shows that our approach is able to integrate and combine successfully different traces from different popular sources into coherent episodes/activities. In addition, we show evidence that our approach does augment user's memory of their past actions, thereby forms a powerful retrospective memory aid. Location: Virtual Committee: Prof. Amelie Marian (Chair)
Prof. Alexander Borgida
Prof. Gerard De Melo
Prof. Chirag Shah, University of Washington
|
| Start Date: 12 Jul 2022; Start Time: 10:00AM - Title: HARDWARE-SOFTWARE TECHNIQUES FOR ACCELERATING SPARSE COMPUTATION Bio: Speaker: Abstract: Linear algebra kernels are widely used in various fields such as machine learning, data science, physical science, and graph analysis. Many of these applications work with sparse data (i.e., only a small fraction of data is non-zero). Sparse data are often stored in a compressed format (i.e., sparse format) that stores only the non-zero elements with additional metadata to identify where the non-zero elements are located. Using compressed formats eliminates the need to store and process zeros, making the storage and computation of sparse kernels more efficient.General purpose architectures, such as CPUs and GPUs, are not able to deliver the same performance for sparse linear algebra kernels as they do for dense versions. First, accessing non-zero elements in sparse format introduces many indirect and irregular memory accesses incompatible with SIMD and caching mechanisms used by CPUs and GPUs. In addition, Dennard scaling is obsolete and Moore's law is slowing down, ending the era in which general-purpose architectures become faster and more energy efficient transparently. This has led to a plethora of research into developing specialized hardware, such as FPGAs and ASICs to improve the performance and energy efficiency of these sparse kernels. A key strategy for the specialized hardware is to customize the sparse format (i.e., storage) according to the operation memory access pattern, the pattern of non-zero elements in the input (i.e., sparsity pattern), and the underlying hardware structures. This approach is effective if the operations and input sparsity patterns do not change. However, applications often perform various operations on sparse data. Additionally, the sparse inputs may frequently change for each execution, and each input may have a different sparsity pattern. When this happens, the performance of specialized hardware degrades because a reformatting step is required to convert the data into a format that is compatible with the hardware. The data reformatting can be expensive when it cannot be overlapped with the computation on the hardware or amortized over multiple application executions with the same input data. This dissertation presents a few hardware-software techniques that enhance the performance and energy efficiency of some of the most important sparse problems, including sparse matrix-vector multiplication (SpMV), sparse general matrix-matrix multiplication (SpGEMM), and sparse convolutional neural networks (CNNs). The key insight of our method is to use the software to reformat the sparse data into a hardware-friendly format, allowing the hardware to perform the computation with a high degree of parallelism. The software improves design flexibility by supporting multiple sparse formats, and the hardware improves performance and energy efficiency. We applied these hardware-software techniques to SpMV, SpGEMM, and sparse CNNs. These problems have different characteristics, such as different input densities and distinct input sparsity pattern features. The contribution of this dissertation can be summarized as follows. First, we present a synergistic CPU-FPGA system to accelerate SpMV and SpGEMM kernels. In our proposed design, the CPU reorganizes sparse data into a format suitable for the FPGA, and the FPGA computes with high parallelism using the preprocessed data. We develop an intermediate representation that allows the software to communicate regularized data and scheduling decisions to the FPGA. Besides, most of the CPU and FPGA execution are overlapped. Our approach can effectively handle sparse kernels with low input densities and sparsity patterns varying for each sparse input. Second, we present a hardware accelerator for sparse CNN inference tasks. We formulate the convolution operation as general matrix-matrix multiplication (GEMM) using an image to column (IM2COL) transformation. With a dynamically reconfigurable GEMM and a novel IM2COL hardware unit, our design can support various layers in CNNs with high performance. Besides, our design exploits sparsity in both weights and feature maps. We use the software to perform group-wise pruning followed by a preprocessing step that puts the pruned weights into our hardware-friendly sparse format for efficient and high performance computation. We built an ASIC and an FPGA prototype of our accelerator. Our design is faster and more energy efficient than CPUs and GPUs for sparse CNNs. Location: Virtual Committee: Professor Santosh Nagarakatte (Chair) Professor Richard P. Martin Professor Yipeng Huang Professor Joe Devietti |
| Start Date: 14 Jul 2022; Start Time: 02:00PM - Title: Application of Deep Neural Networks to Automated Analysis of Hatching Lines for Artist Identification Bio: Speaker: Abstract: Automated analysis of art has gained interest in the computer science research community over the past decade, resulting in the onset of computer vision solutions for authentication and attribution of art. Attribution and authentication are two of the most crucial tasks in the domain of art. Several studies have approached these tasks with an emphasis on analyzing paintings in the visual spectrum; however, with a limited scope to a specific artist, limited datasets, or without rigorous evaluations of the robustness of such approaches. Contrary to paintings, the application of computer vision methodologies for authentication and attribution of drawings and prints has been only sparsely explored. In this dissertation, we introduce a novel approach for attribution and authentication of drawings and sketches at the visual spectrum by automated analysis of hatching lines. Hatching is a prevalent technique in drawing and printmaking, often used to introduce toning, shading, or illusion of light and volume. Since artists tend to add hatching lines spontaneously, we hypothesize that these areas of drawings could carry unique physical and unconscious characteristics of the artist, which can be used to identify the artist. We further hypothesize that such unique characteristics can be quantified by a computational model that will enable automatic attribution and authentication. We investigated the application of deep convolutional neural networks for detecting hatching lines and identifying artists based on their hatching characteristics. We conducted several experiments on different sets of drawings and prints by different artists from different eras, using different techniques and with various degrees of complexity. We concluded that we can indeed identify artists based solely on their hatching with an accuracy of 90-100% in most cases. Location: Virtual Committee: Professor Ahmed Elgammal (Chair) Professor Casimir Kulikowski Professor Abdeslam Boularias Professor Lior Shamir (Kansas State University) |
| Start Date: 15 Jul 2022; Start Time: 10:00AM - 12:00PM Title: Non-neuronal Computational Principles for Increased Performance in Brain-inspired Networks Bio: Speaker: Abstract: Understanding the brain’s computational principles holds potential to lead to human-level machine learning and intelligence. Yet, current brain-inspired spiking neural network (SNN) methods assume that only neurons form the basis of computation in the brain, which contradicts increasing biological evidence associating the function of ubiquitous non-neuronal cells, known as astrocytes, to brain cognitive states, memory and learning. Interestingly, these information processing functions are closely linked to near-critical neuronal network dynamics, which are believed to underlie the brain’s impressive computational abilities. Inspired by this emerging data, in this talk I will present our efforts to understand the computational role of astrocytes in the brain and to translate our findings into brain-inspired neuron-astrocyte spiking networks. First, I will present our biologically plausible computational modeling approach to propose that astrocytes learn to detect and signal deviations in brain activity away from critical dynamics. Further, we suggest that astrocytes achieve this function by approximating the information content of their neuronal inputs. Translating these findings into a well established brain-inspired machine learning paradigm known as the liquid state machine (LSM), I will next present our proposed neuron-astrocyte liquid state machine (NALSM). Specifically, the NALSM uses a single astrocyte to organize recurrent (liquid) network dynamics at the critical regime where the LSM computational capacity peaks. We demonstrate that the NALSM achieves state-of-the-art LSM accuracy and does not need data-specific hand-tuning as do comparable LSM methods. Further, we demonstrate that the NALSM achieves comparable performance to fully-connected multilayer spiking neural networks trained via backpropagation, with a top accuracy of 97.61% on MNIST, 97.51% on N-MNIST, and 85.84% on Fashion-MNIST. Overall, our dual pronged approach leads to testable biological hypotheses about brain computation, which are then abstracted away and integrated directly into brain-inspired machine learning methods. Location: Virtual Committee: Dr. Konstantinos Michmizos (Chair), Dr. Casimir Kulikowsk Dr. Dimitris Metaxa Dr. Constantinos Siettos (University of Naples Federico II, Italy) |
| Start Date: 19 Jul 2022; Start Time: 10:00AM - Title: FAST METHODS TO DETECT AND DEBUG NUMERICAL ERRORS WITH SHADOW EXECUTION Bio: Speaker: Abstract: The floating-point (FP) representation uses a finite number of bits to approximate real numbers in computer systems. Due to rounding errors, arithmetic using the FP representation may diverge from arithmetic with real numbers. For primitive FP operations, the rounding error is small and bounded. However, with the sequence of FP operations, rounding errors can accumulate. Such rounding errors can be magnified by certain operations. The magnified error can affect the program's control flow and the output compared to an execution with infinite bits of precision. It is challenging for users to identify such bugs, as in such cases the program does not crash but generates the incorrect output. Without any oracle in high precision, it is hard to differentiate between correct and incorrect output. Detecting such bugs in long-running programs becomes even more challenging. This dissertation proposes a fast, yet precise mechanism to detect and debug numerical errors in long-running programs. This dissertation makes the following contributions. First, we propose a selective shadow execution framework to detect and debug numerical errors. Our idea is to use shadow execution with high-precision computation for comprehensive numerical error detection. The FP variable is shadowed with a high precision value in this approach. On every FP computation, an equivalent high precision computation is performed. If there is a significant difference between FP computation and high precision computation, the error is reported to the user. To debug errors, we maintain the additional information about the instructions. We use additional information to generate the directed acyclic graph (DAG) of instructions showing the error propagation. The DAG of instructions helps the user identify the root cause of the error. Our prototype FPSanitizer for floating-point is an order of magnitude faster than the prior work. Second, we propose a novel technique to run shadow execution in parallel to reduce the overheads further. In our approach, the user specifies parts of the program that need to be debugged. Our compiler creates shadow execution tasks that mirror these specified regions in the original program but perform equivalent high precision computation. To execute the shadow tasks in parallel, we need to break the dependency between them by providing the appropriate memory state and input arguments. Moreover, to correctly detect the numerical errors in the original program, shadow tasks need to follow the same control flow as in the original program. Our key insight is to use FP values computed by the original program to start the shadow task from some arbitrary point. To ensure shadow tasks follow the same control flow as the original program, our compiler updates every branch instruction in the shadow task to use the branch outcomes of the original program. As a result, the original program and shadow tasks execute in a decoupled fashion and communicate via a non-blocking queue. Our prototype PFPSANITIZER is significantly faster than the FPSANITIZER.Finally, we propose an alternative lightweight oracle to reduce the overheads of shadow execution. Although parallel shadow execution effectively reduces overheads, it requires user input. Often, users may not know where the numerical bugs are present. This thesis proposes a fast shadow execution framework, EFTSANITIZER, that uses error-free transformations (EFTs) to detect anddebug numerical bugs comprehensively. EFTs are a set of algorithms that provide a mechanism to capture the rounding error using primitive FP instructions. For certain FP computations rounding error can be represented as an FP value. Based on this observation, EFTs transform FP computation to derive the rounding error. In our approach, we maintain the error with each memory location and compute the error for a sequence of FP operations using error composition with EFTs. EFTSANITIZER provides a trace of instructions that help users isolate and debug the root cause of the errors. In addition, EFTSANITIZER is an order of magnitude faster than FPSANITIZER. Location: Virtual Committee: Professor Santosh Nagarakatte (Chair)
Professor Richard Martin
Professor Mridul Aanjaneya
Professor Sreepathi Pai (University of Rochester) |
| Start Date: 08 Sep 2022; Start Time: 01:00PM - Title: Generative Modeling for Multimodal Data Fusion Bio: Speaker: Abstract: Humans can process multiple perspectives of the world and generate compound information from them to obtain a richer understanding of the world.In machine learning, generative models have been widely used to comprehend the world by learning the latent representation of the given data. Representation learning is a key step in the process of data understanding, where the goal is to distill interpretable factors associated with the data. However, representation learning approaches typically focus on data observed in a single modality, such as text, images, or video. In this dissertation, we develop generative models that can fuse multimodal data in order to have a comprehensive understanding of the tasks. First, we introduce a model that can learn the latent representation given multi-modality data based on the VAE framework. Specifically, we disentangle the latent space into modality-common (shared) and modality-specific (private) spaces. By considering the private latent factors aside from the shared latent factors of all modalities, the proposed model can achieve more precise cross-modal generation or retrieving from one modality to another modality. Moreover, by assuming the missing modality, we demonstrate that our model can be utilized for solving semi-supervised learning or zero-shot learning problems. We further study the importance of perceiving the world from multiple views through trajectory forecasting scenarios. The growing demand for autonomous vehicles is spurring numerous studies of behavior prediction. The existing works underestimate the context such as nearby environment and neighbors while predicting the target's future trajectories. We suggest combining such information, and generate the context-aware future trajectory prediction conditioned on the past trajectory based on the Conditional VAE. We show that the proposed model is able to achieve collision-free predictions from the surrounding environment or neighbors in both real and simulated data. Location: Virtual Committee: Professor Vladimir Pavlovic, Chair Professor Mubbasir Kapadia Professor Dimitris Metaxas Professor Kris Kitani (Carnege Mellon University) |
| Start Date: 12 Sep 2022; Start Time: 10:00AM - 12:00PM Title: Learning Compositional Robotic Manipulation Tasks from Unlabeled Data Bio: Speaker: Abstract: Researchers have been seeking intelligent robotic systems that can accomplish complex tasks autonomously with very little human effort. With the recent progress from both planning algorithms and learning based methods, many low-level primitives can be fulfilled with remarkable quality. This thesis aims at the next step, compositional manipulation tasks. A compositional manipulation task consists of multiple sub-tasks and each sub-tasks can be completed by low-level manipulation, it requires a high-level reasoning to schedule sub-tasks. In order to reduce labeling effort, we concentrate on self-supervised learning methods. To be specific, we study reinforcement learning algorithms and imitation learning algorithms. Reinforcement learning algorithms explore in the environment and find optimal behavior based on rewards. As the only feedback is a reward and there is no direct information about sub-tasks provided, we utilize first causality in training predictive model. Another proposed method is to construct a finite-state machine for describing transitions between sub-tasks and include it in the policy input. Another line of works covered by this thesis are imitation learning methods where the robot is given a collection of human demonstrations rather than the reward. These demonstrations always present a complete task rather than sub-tasks, labels on sub-tasks switching are never provided. We propose a decomposition of the policy, and maximization of demonstration trajectory likelihood based on this decomposition learns the sub-tasks switching autonomously. Finally, we investigate generalizing this manipulation algorithm to unseen objects by removing the requirement of semantic labels on objects. The proposed method describes objects by feature vectors including appearance and shape information, so similar objects share similar features. Location: Virtual Committee: Professor Abdeslam Boularias (Chair) Professor Jingjin Yu Professor Mridul Aanjaneya Professor Oliver Kroemer (Carnegie-Mellon University) |
| Start Date: 12 Sep 2022; Start Time: 10:30AM - Title: A Case for Correctly Rounded Elementary Functions Bio: Santosh Nagarakatte is an Associate Professor and Undergraduate Program Director of Computer Science at Rutgers University. He obtained his PhD from the University of Pennsylvania in 2012. His research interests are in Hardware-Software Interfaces spanning Programming Languages, Compilers, Software Engineering, and Computer Architecture. His group's research has been recognized with the NSF CAREER Award, two IEEE Micro Top Picks Paper Awards (2010 and 2013), five Distinguished Paper Awards (PLDI 2015, ICSE 2016, PLDI 2021, POPL 2022, and CGO 2022), SIGPLAN Research Highlights paper, CACM Research Highlight paper, 2018 ACM SIGPLAN John C Reynolds Outstanding Dissertation Award, Google Research Award, Intel Corporation Gifts, Facebook research award, and 2022 ACM SIGPLAN John C Reynolds Outstanding Dissertation Award. Speaker: Abstract: This talk will provide an overview of the RLIBM project where we are building a collection of correctly rounded elementary functions for multiple representations and rounding modes. Historically, polynomial approximations for elementary functions have been designed by approximating the real value.In contrast, we make a case for approximating the correctly rounded result of an elementary function rather than the real value of an elementary function in the RLIBM project. Once we approximate the correctly rounded result, there is an interval of real values around the correctly rounded result such that producing a real value in this interval rounds to the correct result. This interval is the freedom that the polynomial approximation has for an input, which is larger than the ones with the mini-max approach. Using these intervals, we structure the problem of generating polynomial approximations that produce correctly rounded results for all inputs as a linear programming problem. The results from the RLIBM project makes a strong case for mandating correctly rounded results with any representation that has fewer than or equal to 32-bits. Read more about the RLIBM project at https://people.cs.rutgers.edu/~sn349/rlibm/ Location: CoRE 301 + Virtual Committee: |
| Start Date: 15 Sep 2022; Start Time: 09:00AM - 11:00AM Title: Multi-Dimensional Federated Learning In Recommender Systems Bio: Speaker: Abstract: A wide range of web services like e-commerce, job-searching, and target advertising heavily rely on recommender systems that finds products of interest to fulfills users' diverse and complicated demands. To better model the user preferences and provide satisfactory recommendations, there has been an increasing amount of research focusing on constructing more accurate and complete user representations that exploit the user profile and behavior history. Inevitably, this motion would induce privacy risks for users. This natural conflict between user privacy and recommendation accuracy has drawn lots of attention in recent years. Among these solutions, the most widely studied and verified method is the federated learning techniques. The general idea behind federated learning is maintaining users' critical data on the edge devices (e.g. mobile phones) which communicate only the model parameters to the central server.However, the standard federated learning system is designed for a single task with a simple learning objective. In reality, a user typically interacts with various applications everyday with heterogeneous intentions. In this defense, I will discuss the extended federated learning in a more complex but practical multi-objective setting, where multiple federated learning agents collaborate in an environment with multiple central servers and a number of distributed edge devices. Formally, this machine learning problem is regarded as a multi-objective federated optimization problem, and I identify several main challenges including the dimensional heterogeneity, conflicting objectives, multi-dimensional communication overhead. Then, I will illustrate the general solution framework and illustrate several techniques that could solve the aforementioned challenges. Location: Virtual Committee: Professor Amelie Marian (Chair) Professor Yongfeng Zhang (Co-advisor) Professor Hao Wang Professor Qingyao Ai (University of Utah) |
| Start Date: 19 Sep 2022; Start Time: 10:00AM - Title: Generative Model and Latent Space Based Medical Image Analysis and Applications Bio: Speaker: Abstract: The generative models have gained much attention in the computer vision community in many cases, such as domain adaptation, data augmentation, and image-to-image translation. These properties have been rapidly applied in some medical domains, but the generative model's potential is not fully explored. This dissertation studies new perspectives on the distributed generative models and latent space manipulation to address some of the most critical medical tasks: medical image private data sharing and cardiac motion analysis. First, we work on asynchronized distributed GAN(AsynDGAN) paradigm to learn the distribution across several private medical data centers and adopt the well-trained generator as a medical data provider for the future use of the downstream tasks. Further, I work on some real scenarios under the continuous learning(Life-long learning) settings with the distributed GAN with temporary discriminators(TDGAN). Such a method could prevent the model from catastrophic forgetting when continuously learning new incoming data. The multi-modality and missing-modality settings are also systematically analyzed. By using a multi-modality adaptive learning model and network(Modality Bank), the Modality Bank could auto-complete the missing modalities and generate multiple modality images simultaneously. We demonstrate that the AsynDGAN-related techniques could secure medical privacy while fully using these private data for machine learning applications. Secondly, the dissertation presents a framework for joint 2D cardiac segmentation and 3D volume reconstruction via a structure-specific generative method(DeepRecon). I will present the end-to-end latent-space-based framework that generates multiple clinically essential outcomes, including accurate image segmentation, synthetic high-resolution 3D image, and 3D reconstructed volume. Experimental results demonstrate the effectiveness of our approach on numerous fronts, including 2D segmentation, 3D reconstruction, and downstream 4D motion pattern adaption performance. And the motion adaptation method provides a unique tool to help cardiologists analyze cardiac motion functional differences between various cases. Overall, the approaches demonstrate the importance of the generative models for the newly emerging medical analysis domains for 3D reconstruction, motion analysis, and privacy data sharing. Location: Virtual Committee: Professor Dimitris N. Metaxas (Chair) Professor Konstantinos P. Michmizos Professor Karl Stratos Professor Xiaolei Huang(Penn State University) |
| Start Date: 19 Sep 2022; Start Time: 10:30AM - Title: Towards Closing the Perception-Planning and Sim2Real Gaps in Robotics Bio: Kostas Bekris is an Associate Professor of Computer Science at Rutgers University in New Jersey. He is working in algorithmic robotics, where his group is developing algorithms for robot planning, learning and perception especially for robot manipulation and multi-robot problems. Applications include logistics and manufacturing with a focus on taking advantage of novel soft, adaptive mechanisms. His research has been supported by NSF, DHS, DOD and NASA, including a NASA Early Career Faculty award. He received his Ph.D in Computer Science from Rice University under the guidance of Prof. Lydia Kavraki. Kostas is Program Chair for the “Robotics: Science and Systems” (RSS) 2023 conference. Speaker: Abstract: Robotics is at the point where we can deploy complete systems across applications, such as logistics, service and field robotics. There are still critical gaps, however, that limit the adaptability, robustness and safety of robots, which lie at: (a) the interface of domains, such as perception, planning and learning, that must be viewed holistically in robotics, and (b) the sim2real gap, i.e., the deviation between internal models of robots’ AI and the real world. This talk will first describe efforts in tighter integration of perception and planning for vision-driven robot manipulation. We have developed high-fidelity, high-frequency tracking of rigid bodies’ 6D poses - without using CAD models or cumbersome human annotations - by utilizing progress both in deep learning and pose graph optimization. These solutions together with appropriate shared representations, tighter closed-loop operation and compliant mechanisms are unblocking the deployment of full-stack robot manipulation systems. This talk will provide examples of robust robotic packing, assembly under tight tolerances as well as constrained placement given a single demonstration that generalizes across an object category. The talk’s second part is motivated by tensegrity robots, which combine rigid and soft elements, to achieve safety and adaptability. They also complicate, however, modeling and control given their high-dimensionality and complex dynamics. This sim2real gap of analytical models has motivated us to look into reinforcement learning (RL) for controlling robot tensegrities, which allowed the development of new skills for them. RL applicability is limited, however, due to its high data requirements. Training RL in simulation is promising but is blocked again by the sim2real gap. For this reason, we are developing differential engines for tensegrity robots that reason about first-principles so as to be trained with few example trajectories from the real robot. They provide accurate-enough simulations to train a controller that is directly transferrable back to the real system. We report our first success in such a real2sim2real transfer for a 3-bar tensegrity robot. The talk will conclude with a brief discussion on how closing these gaps empowers the next step of developing robots that are socially cognizant and can be safely integrated into our society. Location: CoRE 301 + Virtual Committee: |
| Start Date: 20 Sep 2022; Start Time: 11:30AM - 01:30PM Title: Projections, Extractors, and Streaming Lower Bounds Bio: Speaker: Abstract: See Above Location: Virtual Committee: Professor Swastik Kopparty, Chair (University of Toronto) Professor Eric Allender Professor Sepehr Assadi Professor Huacheng Yu (Princeton University) |
| Start Date: 23 Sep 2022; Start Time: 11:00AM - Title: Closing the Reality Gap for Controlling Tensegrity Robots via Differentiable Physics Engines Bio: Speaker: Abstract: See above Location: Virtual Committee: Professor Kostas Bekris (Chair) Professor Mridul Aanjaneya Professor Abdeslam Boularias Professor Andrew Sabelhaus (Boston University) |
| Start Date: 04 Oct 2022; Start Time: 11:00AM - Title: Hybrid robotics and implicit learning Bio: Michael Posa is an Assistant Professor in Mechanical Engineering and Applied Mechanics at the University of Pennsylvania. He leads the Dynamic Autonomy and Intelligent Robotics (DAIR) lab, a group within the Penn GRASP laboratory. His group focuses on developing computationally tractable algorithms to enable robots to operate both dynamically and safely as they quickly maneuver through and interact with their environments, with applications including legged locomotion and manipulation. Michael received his Ph.D. in Electrical Engineering and Computer Science from MIT in 2017, where, among his other research, he spent time on the MIT DARPA Robotics Challenge team. He received his B.S. in Mechanical Engineering from Stanford University in 2007. Before his doctoral studies, he worked as an engineer at Vecna Robotics in Cambridge, Massachusetts, designing control algorithms for the BEAR humanoid robot. He has received the Best Paper award at Hybrid Systems: Computation and Control (HSCC) and been finalist awards at ICRA and IEEE Humanoids. He has also received Google Faculty Research Award in 2019 and the Young Faculty Researcher Award from the Toyota Research Institute in 2021. Speaker: Abstract: Machine learning has shown incredible promise in robotics, with some notable recent demonstrations in manipulation and sim2real transfer. These results, however, require either an accurate a priori model (for simulation) or a large amount of data. In contrast, my lab is focused on enabling robots to enter novel environments and then, with minimal time to gather information, accomplish complex tasks. In this talk, I will argue that the hybrid or contact-driven nature of real-world robotics, where a robot must safely and quickly interact with objects, drives this high data requirement. In particular, the inductive biases inherent in standard learning methods fundamentally clash with the non-differentiable physics of contact-rich robotics. Focusing on model learning, or system identification, I will show both empirical and theoretical results which demonstrate that contact stiffness leads to poor training and generalization, leading to some healthy skepticism of simulation experiments trained on artificially soft environments. Fortunately, implicit learning formulations, which embed convex optimization problems, can dramatically reshape the optimization landscape for these stiff problems. By carefully reasoning about the roles of stiffness and discontinuity, and integrating non-smooth structures, we demonstrate dramatically improved learning performance. Within this family of approaches, ContactNets accurately identifies the geometry and dynamics of a six-sided cube bouncing, sliding, and rolling across a surface from only a handful of sample trajectories. Similarly, a piecewise-affine hybrid system with thousands of modes can be identified purely from state transitions. I'll also discuss how these learned models can be deployed for control via recent results in real-time, multi-contact MPC. Location: 1 Spring street, New Brunswick, NJ – Room 403 + Virtual Committee: |
| Start Date: 19 Oct 2022; Start Time: 05:30PM - Title: SAViR-T: Spatially Attentive Visual Reasoning with Transformers Bio: Speaker: Abstract: Visual Reasoning (VR) operates as a way to measure machine intelligence, by employing previously gained knowledge in new settings. Specifically, in VR, we aim to extract and identify task-relevant information from images. For example, in Raven's Progressive Matrices (RPMs), an instance of VR, we are given an incomplete 3x3 image puzzle. We should find the governing rules that generated the puzzle in order to solve it. In this talk, we will explore the importance of localized spatial information for the solution of RPM puzzles. Our proposed model SAViR-T considers explicit spatial semantics of visual elements within each image in the puzzle, encoded as spatio-visual tokens, and learns the intra-image as well as the inter-image token dependencies. Token-wise relationships, modeled through a transformer-based SAViR-T architecture, followed by a reasoning module are used to extract the underlying rule representations between the rows of the RPM. We use these relation representations to complete the puzzle. Finally, to demonstrate the efficacy of our approach we performed extensive experiments across both synthetic datasets, including RAVEN, I-RAVEN, RAVEN-FAIR, and the natural image-based "V-PROM". Location: Virtual Committee: Professor Vladimir Pavlovic Professor Srinivas Narayana Ganapathy Professor Hao Wang Professor Yongfeng Zhang |
| Start Date: 21 Oct 2022; Start Time: 10:00AM - Title: Learning-Based Robot Control from Vision: Formal Guarantees and Fundamental Limits Bio: Anirudha Majumdar is an Assistant Professor at Princeton University in the Mechanical and Aerospace Engineering (MAE) department, and Associated Faculty in the Computer Science department. He also holds a part-time position as a Visiting Research Scientist at the Google AI Lab in Princeton. He received a Ph.D. in Electrical Engineering and Computer Science from the Massachusetts Institute of Technology in 2016, and a B.S.E. in Mechanical Engineering and Mathematics from the University of Pennsylvania in 2011. Subsequently, he was a postdoctoral scholar at Stanford University from 2016 to 2017 at the Autonomous Systems Lab in the Aeronautics and Astronautics department. He is a recipient of the ONR YIP award, the NSF CAREER award, the Google Faculty Research Award (twice), the Amazon Research Award (twice), the Young Faculty Researcher Award from the Toyota Research Institute, the Best Conference Paper Award at the International Conference on Robotics and Automation (ICRA), the Paper of the Year Award from the International Journal of Robotics Research (IJRR), the Alfred Rheinstein Faculty Award (Princeton), and the Excellence in Teaching Award from Princeton’s School of Engineering and Applied Science. Speaker: Abstract: The ability of machine learning techniques to process rich sensory inputs such as vision makes them highly appealing for use in robotic systems (e.g., micro aerial vehicles and robotic manipulators). However, the increasing adoption of learning-based components in the robotics perception and control pipeline poses an important challenge: how can we guarantee the safety and performance of such systems? As an example, consider a micro aerial vehicle that learns to navigate using a thousand different obstacle environments or a robotic manipulator that learns to grasp using a million objects in a dataset. How likely are these systems to remain safe and perform well on a novel (i.e., previously unseen) environment or object? How can we learn control policies for robotic systems that provably generalize to environments that our robot has not previously encountered? Unfortunately, existing approaches either do not provide such guarantees or do so only under very restrictive assumptions.In this talk, I will present our group’s work on developing a framework for learning control policies for robotic systems with formal guarantees on generalization to novel environments. The key technical insight is to leverage and extend powerful techniques from generalization theory in theoretical machine learning. We apply our techniques on problems including vision-based navigation and manipulation in order to demonstrate the ability to provide strong generalization guarantees on robotic systems with complicated (e.g., nonlinear/hybrid) dynamics, rich sensory inputs (e.g., RGB-D), and neural network-based control policies. I will also present recent work aimed at understanding fundamental limits on safety and performance imposed by a robot’s (imperfect) sensors. Location: Room 403, 1 Spring street, New Brunswick, NJ + Virtual Committee: |
| Start Date: 21 Oct 2022; Start Time: 05:00PM - 07:00PM Title: Complete and Efficient Prehensile Rearrangement in Confined Spaces under Kinematic Constraints Bio: Speaker: Abstract: Rearranging objects in confined spaces has broad applications, such as rearranging products in grocery shelves (e.g., restocking), retrieving food from a packed refrigerator and assembling a product out of individual components. Meanwhile solving such problems in confined spaces is challenging as no top-down grasps are available for approaching the objects, which simplify rearrangement tasks on tabletops. As a result, these problems involve challenging kinematic and geometric constraints, which include both robot-to-object and object-to-object interactions.This thesis is motivated by this domain and proposes task and motion planning algorithms that result in a high-quality sequence of robot motions, which allow to successfully complete rearrangement tasks in confined spaces without undesirable collisions. Specifically, this work introduces efficient monotone solvers for solving monotone problems, i.e., those that can be solved by moving each object at most once, by significantly pruning the search space of possible solutions. In this process, it sidesteps expensive computational operations, while maintaining desirable completeness guarantees. This thesis progresses in incorporating the proposed monotone solvers to develop probabilistically complete non-monotone solvers, which are capable of solving harder instances quickly with fewer buffer locations, i.e., intermediate placements for objects needed during the rearrangement process. This work also provides improved motion planning primitives for rearrangement to speed up online motion planning resolution. The combination of these algorithmic improvements allows for increased feasibility, efficiency and quality of solutions. Finally, this work demonstrates the applicability of the proposed methods via a proof-of-concept real robotic rearrangement system, which integrates visual input and the developed task and motion planning methods. Location: Room 203, 1 Spring Street, New Brunswick, 08901 Committee: Professor Kostas Bekris (Chair) |
| Start Date: 02 Nov 2022; Start Time: 12:00PM - 01:00PM Title: A Manifold View of Adversarial Risk Bio: Speaker: Abstract: The adversarial risk of a machine learning model has been widely studied. Most previous works assume that the data lies in the whole ambient space. We propose to take a new angle and take the manifold assumption into consideration. Assuming data lies in a manifold, we investigate two new types of adversarial risk, the normal adversarial risk due to perturbation along normal direction, and the in-manifold adversarial risk due to perturbation within the manifold. We prove that the classic adversarial risk can be bounded from both sides using the normal and in-manifold adversarial risks. We also show with a surprisingly pessimistic case that the standard adversarial risk can be nonzero even when both normal and in-manifold risks are zero. We finalize the paper with empirical studies supporting our theoretical results. Our results suggest the possibility of improving the robustness of a classifier by only focusing on the normal adversarial risk. Location: https://rutgers.zoom.us/j/93084168396?pwd=OFZZeHJOYWlydEEyRDZmdW40aitQdz09 Committee: Prof. Dimitris Metaxas Dr. Hao Wang Prof. Konstantinos Michmizos. Prof. Xiong Fan |
| Start Date: 04 Nov 2022; Start Time: 11:00AM - 12:15PM Title: Foundations of Transaction Fee Mechanism Design Bio: Elaine Shi is an Associate Professor at Carnegie Mellon University. Her research interests include cryptography, algorithms, and foundations of blockchains. Prior to CMU, she taught at the University of Maryland and Cornell University. She is a recipient of the Packard Fellowship, the Sloan Fellowship, the ONR YIP Award, the NSA best scientific cybersecurity paper award, and various other best paper awards. Speaker: Abstract: Space in a blockchain is a scarce resource. Cryptocurrencies today use auctions to decide which transactions get confirmed in the block. Intriguingly, classical auctions fail in such a decentralized environment, since even the auctioneer can be a strategic player. For example, the second-price auction is a golden standard in classical mechanism design. It fails, however, in the blockchain environment since the miner can easily inject a bid that is epsilon smaller than the k-th price where k is the block size. Moreover, the miner and users can also collude through the smart contract mechanisms available in modern cryptocurrencies. I will talk about a new foundation for mechanism design in a decentralized environment. I will prove an impossibility result which rules out the existence of a dream transaction fee mechanism that incentivizes honest behavior for the user, the miner, and a miner-user coalition at the same time. I will then argue why the prior modeling choices are too draconian, and how we can overcome this lower bound by capturing hidden costs pertaining to certain deviations. Location: Core 301 Committee: |
| Start Date: 09 Nov 2022; Start Time: 08:00AM - 05:00PM Title: Towards Real-time Scene-level Tracking and Reconstruction Bio: Speaker: Abstract: Real-time perception is a crucial component of modern robotic manipulation systems. Recent progress in manipulation area has demonstrated that given the geometry model and 6D-pose trajectory of a manipulated object during an expert demonstration, a robot can quickly learn complex and contact-rich manipulation skills. However, a system that can simultaneously provide a geometry model and 6-D pose is notoriously hard to build. In this talk, we present the first real-time system, STAR-no-prior capable of tracking and reconstructing, individually, every visible object in a given scene, without any form of prior on the rigidness of the objects, texture existence, or object category. Despite its strong performance, STAR-no-prior is limited to its multi-camera setting. To compensate for this, we further develop Mono-STAR, the first real-time mono-camera 3D reconstruction system that simultaneously supports semantic fusion, fast motion tracking, non-rigid object deformation, and topological change under a unified framework. Location: https://rutgers.zoom.us/j/92713978624?pwd=UXVyeXBibkdObHJkUGdScmVOMmlPUT09 Meeting ID: 927 1397 8624 Password: 981315 Committee: Abdeslam Boularias Kostas Bekris Mridul Aanjaneya Dong Deng |
| Start Date: 11 Nov 2022; Start Time: 02:30PM - 04:00PM Title: Space Optimal Matching and Vertex Cover in dynamic streams Bio: Speaker: Abstract: We present algorithms for maximum matching and vertex cover in dynamic (insertion-deletions) streams with asymptotically optimal space complexity: for any n-vertex graph, our algorithms with high probability output an α-approximate matching in a single pass using O(n^2/α^3) bits of space and an α-approximate vertex cover using O(n^2/α^2) bits of space.A long line of work on the dynamic streaming matching problem has reduced the gap between space upper and lower bounds first to n^o(1) factors [Assadi-Khanna-Li-Yaroslavtsev; SODA 2016] and subsequently to polylog(n) factors [Dark-Konrad; CCC 2020]. [Dark-Konrad; CCC 2020] also gave upper and lower bounds for streaming vertex cover that were only a polylog(n) factor apart. Our upper bounds now match the DarkKonrad lower bounds up to O(1) factors, thus completing this research direction.Our approach consists of two main steps: we first (provably) identify a family of graphs, similar to the instances used in prior work to establish the lower bounds for this problem, as the only “hard” instances to focus on. These graphs include an induced subgraph which is both sparse and contains a large matching. We then design dynamic streaming algorithms for this family of graphs which is more efficient than prior work. The keys to this efficiency are novel sketching methods, which bypass the typical loss of polylog (n)-factors in space compared to standard L0-sampling primitives, and can be of independent interest in designing optimal algorithms for other streaming problems. Location: Zoom information: https://rutgers.zoom.us/j/91956380264?pwd=c1I0bFk3dkhwalVQMjlmZk1MY3ZrUT09 Meeting ID: 919 5638 0264 P Committee: Prof. Aaron Bernstein Prof. Jie Gao Prof. Sepehr Assadi Prof. Yipeng Huang |
| Start Date: 17 Nov 2022; Start Time: 10:30AM - Title: Unstructured Data Management at Scale for Machine Learning Bio: Dong Deng is an assistant professor in the Computer Science Department at Rutgers University. His research interests include large-scale data management, data science, data curation, and database systems. Before joining Rutgers, he was a postdoc in the Database Group at MIT, where he worked with Mike Stonebraker and Sam Madden on data curation systems. He received his Ph.D. from Tsinghua University with honors. He has published over 30 research papers in top data management venues, mainly SIGMOD, VLDB, and ICDE. Based on Google Scholar, his publications have attracted over 2000 citations. His research is supported by a couple of NSF awards. He regularly serves the PC committee of various data management, data mining, and information retrieval conferences. He also serves on the organization committees of several data management conferences. Speaker: Abstract: A clear trend in machine learning is that models become larger and larger and more and more training data is used. This is especially true for foundation models, such as GPT-3 and DALL-E 2. In this talk, I will discuss a few unstructured data management problems arising in machine learning. First, recent studies show large language models (LLMs) unintendedly memorize part of the training data, which brings significant privacy risks. These studies mostly focus on the exact duplicates. However, how many texts generated by LLMs have near-duplicate sequences in the training data? Do sequences with more near-duplicates in the training data more likely to be memorized by LLMs? To answer these questions, I will introduce a series of works (SIGMOD-21 and SIGMOD-22a) from my group that enables efficient near-duplicate sequence search in terabytes of LLM training texts on a single machine. A major challenge here is the number of sequences in a text is quadratic to the text length. Second, real-world objects such as images and texts can be represented as dense vectors. I will briefly introduce our work on large-scale vector data management, a project funded by NSF IIS. Finally, I will conclude the talk by outlining a few near-future works to be conducted in my group. Location: CoRE 301 Committee: |
| Start Date: 22 Nov 2022; Start Time: 09:30AM - 12:30PM Title: Fine-grained Air Quality Nowcasting Bio: Speaker: Abstract: Prolonged exposure to air pollution is a health hazard. Users could reduce pollution exposure by accessing an accurate air quality information stream. The current air quality information stream is coarse-grained (measurements spatially and temporally few and far between) and does not reflect the localized air quality variations that a fine-grained (spatially and temporally dense measurements) information stream could. This dissertation aims to fill this need by proposing new sensing and modeling recipes to achieve fine-grained information streams with pollution inventory. In the first part of the dissertation, we present two mobile sensing platforms for fine-grained real-time pollution measurements - a portable sensing platform deployable on public transportation infrastructure and a personal sensing device that can create a social pollution sensing network. We assemble mobile sensing platforms and deploy them on Rutgers campus buses for evaluation. We conclude that mobile sensing platforms deployed on public transportation infrastructure can help collect fine-grained pollution measurements. We also propose a new neural network architecture - "InsideOut," to infer Carbon Monoxide measurements outdoors based on in-vehicle Carbon Monoxide measurements collected by users monitoring their personal space, thus contributing to the fine-grained pollution inventory outdoors. In the second part of the dissertation, we propose "X-PoSuRe" - a neural network-based regression model for pollution super-resolution trained to infer fine-grained pollution information from coarse-grained pollution measurements akin to image super-resolution, where a neural network model creates high-resolution images from low-resolution images. The X-PoSuRe model uses Nitrogen Dioxide measurements and other air quality covariates for pollution super-resolution. The proposed X-PoSuRe model provides a promising new and novel method for pollution super-resolution from existing low-resolution data sources without the need for deploying expensive equipment over a large area. We further extend this model to infer fine-grained measurements for other pollutants.In the third and final step, we evaluate the benefits of a fine-grained pollution inventory by demonstrating on a neighborhood scale that a significant reduction in pollution exposure can be achieved by choosing a healthy way instead of the shortest or quickest way.Overall, our work pushes the frontiers of inference models in modeling and inferring a hard-to-measure entity using its easily measurable covariates. Location: CoRE 305 Committee:
|
| Start Date: 22 Nov 2022; Start Time: 10:45AM - Title: Bayesian Deep Learning: From Single-Domain Reasoning to Infinite-Domain Adaptation Bio: Hao Wang is currently an Assistant Professor in the Department of Computer Science at Rutgers University. Previously he was a Postdoctoral Associate at the Computer Science & Artificial Intelligence Lab (CSAIL) of MIT, working with Dina Katabi and Tommi Jaakkola. He received his PhD degree from the Hong Kong University of Science and Technology, as the sole recipient of the School of Engineering PhD Research Excellence Award in 2017. He has been a visiting researcher in the Machine Learning Department of Carnegie Mellon University. His research focuses on statistical machine learning, deep learning, and data mining, with broad applications on recommender systems, healthcare, user profiling, social network analysis, text mining, etc. His research was recognized and supported by the Microsoft Fellowship in Asia, the Baidu Research Fellowship, the Amazon Faculty Research Award, and the National Science Foundation. Speaker: Abstract: While perception tasks such as visual object recognition and text understanding play an important role in human intelligence, the subsequent tasks that involve inference, reasoning, and planning require an even higher level of intelligence. The past few years have seen major advances in many perception tasks using deep learning models. In terms of higher-level inference, however, probabilistic graphical models, with their ability to expressively describe properties of variables and various probabilistic relations among variables, are still more powerful and flexible. To achieve integrated intelligence that involves both perception and inference, we have been exploring along a research direction, which we call Bayesian deep learning, to tightly integrate deep learning and Bayesian models within a principled probabilistic framework. In this talk, I will present the proposed unified framework and some of our recent work on Bayesian deep learning with various applications including recommendation, social network analysis, interpretable healthcare, domain adaptation, and representation learning. Location: CoRE 301 Committee: |
| Start Date: 28 Nov 2022; Start Time: 10:30AM - 12:00PM Title: Self-Supervised Object Understanding for Robot Perception and Manipulation Bio: Speaker: Abstract: : To be deployed in indoor human environments, autonomous robots should be able to reason about unknown objects so as to manipulate them safely and efficiently. It is, however, challenging because household objects can be arbitrary with various textures and geometries. And it is not feasible to manually annotate or train a model for every single object or even at the category level. We aim to let robots understand objects from their own experiences, either through passively observing objects from different viewpoints and locations, or actively manipulating them. In particular, we designed algorithms to automatically find patterns of rigid objects across different scenes even in clutter, and use the pseudo labels to train a segmentation model via contrastive learning. We also built a robotic system to manipulate objects without access to mesh models. The target object is tracked and reconstructed during the manipulation process. Its texture information and reconstructed model are saved in a memory bank to speed up future manipulation of the same object. Through a lifelong process, the robot becomes more capable over time as it observes and manipulates more objects. In the future, we aim to extend the current research to allow robots to robustly handle objects in more complex scenes. Location: 1 Spring Street, Room 204 Committee: Prof. Kostas Bekris (advisor) Prof. Abdeslam Boularias Prof. Kristin Dana Prof. Sepehr Assadi |
| Start Date: 05 Dec 2022; Start Time: 10:30AM - Title: From Deep Learning to Program Learning Bio: He Zhu is an assistant professor in the Department of Computer Science at Rutgers University. He was a research scientist at Galois, Inc. He received his Ph.D. degree from Purdue University. His work spans programming languages, formal methods, and machine learning. He is currently interested in building intelligent program learning systems that tightly integrate deep learning and program synthesis and that can be formally verified. Dr. Zhu received two distinguished paper awards from the prestigious ACM SIGPLAN conference on Programming Language Design and Implementation. His research was supported by the National Science Foundation and the Defense Advanced Research Projects Agency. Speaker: Abstract: Deep learning has led to encouraging successes in many challenging tasks. However, a deep neural net inherently lacks interpretability due to the difficulty of identifying how the model's learned logic relates to its complex network structure. It is difficult to debug and reason about deep neural nets at the same level developers manage conventional software systems. Program-guided models (i.e. neurosymbolic programs) have recently attracted much interest due to their interpretability and compositionality. Yet, synthesizing programs requires optimizing over a combinatorial, non-differentiable, and rapidly exploded space of program structures. In this talk, I will present our recent efforts on enabling human-readable, domain-specific programs as an efficient learning representation. Powered by novel program synthesis algorithms, our method jointly optimizes program structures and program parameters. As a step toward trustworthy learning, it adapts formal methods typically designed for traditional human-written software systems to provide formal correctness guarantees to program-guided models. Experiment results over application domains such as behavior classification and reinforcement learning demonstrate that our algorithms excel in discovering optimal programs that are highly interpretable and verifiable. Location: CoRE 301 + Virtual Committee: |
| Start Date: 06 Dec 2022; Start Time: 10:30AM - Title: Contrastive Self-Supervised Learning and Deep Pre-trained Language Models for Entity Resolution Bio: Speaker: Abstract: Entity Resolution (ER) is a field of study dedicated to finding items that belong to the same entity, and is an essential problem in NLP and data integration and preparation (DI&P). We propose Sudowoodo, a multi-purpose DI&P framework based on contrastive representation learning and deep pre-trained language models. Sudowoodo features a unified, matching-based problem definition capturing a wide range of DI&P tasks including Entity Resolution (ER) in data integration, error correction in data cleaning, semantic type detection in data discovery, and more. Contrastive learning enables Sudowoodo to learn similarity-aware data representations from a large corpus of data items (e.g., entity entries, table columns) without using any labels. The learned representations can later be either directly used or facilitate fine-tuning with only a few labels to support the ER task. Our experiment results show that Sudowoodo achieves multiple state-of-the-art results on different levels of supervision and outperforms previous best specialized blocking or matching solutions for ER. Sudowoodo also achieves promising results in data cleaning and column matching tasks showing its versatility in DI&P applications. For the blocking step of ER, we propose Neural Locality Sensitive Hashing Blocking (NLSHBlock), which is based on pre-trained language models and fine-tuned with a novel LSH-inspired loss function. NLSHBlock out-performs existing methods on a wide range of datasets. Location: Virtual-The zoom link is https://rutgers.zoom.us/j/93769571337?pwd=NC9nR3RVeDJzcHpUZHl1ZWExK0Jndz09 Committee: Dr. Yongfeng Zhang (advisor) Dr. Hao Wang Dr. Dong Deng Dr. Kostas Bekris (external) |
| Start Date: 12 Dec 2022; Start Time: 10:30AM - Title: Transparent Computing in the AI Era Bio: Shiqing Ma is an Assistant Professor in the Department of Computer Science at Rutgers University, the state university of New Jersey. He received his Ph.D. in Computer Science from Purdue University in 2019 and B.E. from Shanghai Jiao Tong University. His research focuses on program analysis, software and system security, adversarial machine learning, and software engineering. His work on operating system transparency received Distinguished Paper Awards from NDSS 2016 and USENIX Security 2017, and his work on adversarial machine learning received Best Student Paper Awards from ECCV ARWO 2022 and ICTAI 2022. Speaker: Abstract: Recent advances in artificial intelligence (AI) have shifted how modern computing systems work, raising new challenges and opportunities for transparent computing. On the one hand, many AI systems are black boxes and have dense connections among their computing units, which makes existing techniques like dependency analysis fail. Such a new computing system calls for new methods to improve its transparency to defend against attacks against AI-powered systems, such as Trojan attacks. On the other hand, it provides a brand-new computation abstraction, which features data-driven computation-heavy applications. It potentially enables new transparent computing applications, typically involving large-scale data processing. In this talk, I will present my work in these two directions. Specifically, I will discuss the challenges in analyzing the deep neural network for security inspection and introduce our novel approach to examining Trojan behaviors. Later, I will talk about AI can help increase the information entropy of large security audit logs to enable efficient lossless compressed storage. Location: CoRE 301 + Virtual Committee: |
| Start Date: 13 Dec 2022; Start Time: 09:00AM - Title: Generative Video Transformer: Can Objects be the Words? Bio: Speaker: Abstract: Transformers have been successful for many natural language processing tasks. However, applying transformers to the video domain for tasks such as long-term video generation and scene understanding has remained elusive due to the high computational complexity and the lack of natural tokenization. In this paper, we propose the Object-Centric Video Transformer (OCVT) which utilizes an object-centric approach for decomposing scenes into tokens suitable for use in a generative video transformer. By factoring the video into objects, our fully unsupervised model is able to learn complex spatio-temporal dynamics of multiple interacting objects in a scene and generate future frames of the video. Our model is also significantly more memory-efficient than pixel-based models and thus able to train on videos of length up to 70 frames with a single 48GB GPU. We compare our model with previous RNN-based approaches as well as other possible video transformer baselines. We demonstrate OCVT performs well when compared to baselines in generating future frames. OCVT also develops useful representations for video reasoning, achieving start-of-the-art performance on the CATER task. Location: Virtual https://rutgers.zoom.us/j/96462372427?pwd=Tm5vUmxzcGdZZ1RaWGVWUEE0VDNwQT09 Committee: Prof. Sungjin Ahn (Chair) Prof. Hao Wang Prof. Abdeslam Boularias Prof.Yongfeng Zhang(external) |
| Start Date: 21 Dec 2022; Start Time: 03:00PM - 04:30PM Title: Methods for Leveraging Auxiliary Signals for Low-Resource NLP Bio: Speaker: Abstract: There is a growing need for NLP systems that support low-resource settings, for which task-specific training data may be lacking, while domain-specific corpora is too scarce to build a reliable system. In the past decade, the co-occurrence-based training objectives of methods such as word2vec are first able to offer word-level semantic information for specific domains. Recently, pretrained language model architectures such as BERT have been shown capable of learning monolingual or multilingual representations with self-supervised objectives under a shared vocabulary, simply by combining the input from single or multiple languages.Such representations greatly facilitate low-resource language applications.Still, the success of such cross-domain transfer hinges on how close the involved domains are, with substantial drops observed for some more distant domain pairs, such as English to Korean, Wikipedia text to social media comments. To address this, domain-specific unlabeled corpora is available to serve as the auxiliary signals to enhance low-resource NLP systems. In this dissertation, we present a series of methods for leveraging auxiliary signals. In particular, cross-lingual sentiment embeddings with transfer learning are proposed to improve sentiment analysis. For cross-lingual text classification, we present a self-learning framework to take advantage of unlabeled data. Furthermore, a framework upon data augmentation with adversarial training for cross-lingual NLI is proposed for the low-resource problem from the target domain. Extensive experimental results demonstrate the effectiveness of the proposed methods in achieving better performance across a variety of NLP tasks. Location: Virtual https://rutgers.zoom.us/j/8203272500?pwd=R0hGYkNBQllsaWdKMWN3OEh5V0dpZz09 Committee: Prof Gerard de Melo (Chair) Prof Yongfeng Zhang Prof Karl Stratos Prof Handong Zhao (external Member) |
| Start Date: 08 Feb 2023; Start Time: 02:00PM - 03:30PM Title: Building Efficient Storage Systems for Modern Near Storage Data Processing Bio: Speaker: Abstract: See above Location: CoRE 301 Committee: Professor Sudarsan Kannan (Advisor) Professor Santosh Nagarakatte Professor Srinivas Narayana Ganapathy Professor Hao Wang |
| Start Date: 09 Feb 2023; Start Time: 02:00PM - 03:15PM Title: AI for mathematics Bio: PhD In Mathematics, University of Bonn Worked as a research scientist at Cadence Research Laboratories in Berkeley: design automation of digital circuits, physical design and logic synthesis. Christian is a researcher at Google working on machine learning, AI and computer vision via deep learning. Speaker: Abstract: We give an introduction to recent advances in automating mathematics through automated formalization ("autoformalization") and proof-search using deep learning, specifically transformer-based large language models. Autoformalization is the process of automatically transcribing human-written mathematical texts into computer-verifiable proofs. While most natural language mathematics looks fairly formal to the untrained eye, it can take a great deal of human effort to fully formalize mathematical text using "interactive theorem provers". Recent advances in deep-learning-based language modeling and neural-augmented proof search offer a promising path towards autoformalization and human-level mathematical AI. We present recent advances in this area as well as the challenges ahead. Location: Fiber Optics Material Research Building EHA Committee: |
| Start Date: 09 Feb 2023; Start Time: 04:30PM - 06:30PM Title: Representative Learning Enabled Efficient Provenance Data Storage Bio: Speaker: Abstract: See above Location: CoRE 301 Committee: Professor Shiqing Ma (Advisor) Professor Dong Deng Professor Sudarsun Kannan Professor Abdeslam Boularias |
| Start Date: 09 Feb 2023; Start Time: 05:00PM - 07:00PM Title: Deep Learning-based Biomedical Images Classification and Segmentation from Limited Data Bio: Speaker: Abstract: See above Location: CBIM 22 Committee: Professor Dimitris N. Metaxas (Chair) Professor Vladimir Pavlovic Professor Hao Wang Dr. Leon Axel (NYU) |
| Start Date: 10 Feb 2023; Start Time: 02:00PM - 04:00PM Title: Safe Object Rearrangement in Confined Spaces Under Visibility Constraints Bio: Speaker: Abstract: See above Location: 1 Spring Street - Room 319 Committee: Professor Kostas Bekris (Advisor) Professor Jingjin Yu Professor Jie Gao Professor He Zhu |
| Start Date: 23 Feb 2023; Start Time: 10:30AM - 11:30AM Title: A Hybrid Computing Ecosystem For Practical Quantum Advantage Bio: Gokul Subramanian Ravi is a 2020 NSF CI Fellows postdoctoral scholar at the University of Chicago, mentored by Prof. Fred Chong. His research targets quantum computing architecture and systems, primarily on themes at the intersection of quantum and classical computing. He received his PhD in computer architecture from UW-Madison in 2020 and was advised by Prof. Mikko Lipasti. He was awarded the 2020 Best ECE Dissertation Award from UW-Madison and named a 2019 Rising Star in Computer Architecture. His quantum and classical computing research have resulted in publications at top computer architecture, systems, and engineering venues (such as ASPLOS, ISCA, MICRO, HPCA, TACO, ISLPED, QCE, IISWC), as well as three filed and two granted patents. His co-authored work was recognized as the Best Paper at HPCA 2022 and as a 2023 IEEE Micro Top Picks Honorable Mention. Speaker: Abstract: As quantum computing transforms from lab curiosity to technical reality, we must unlock its full potential to enable meaningful benefits on real-world applications with imperfect quantum technology. Achieving this vision requires computer architects to play a key role, leveraging classical computing principles to build and facilitate a hybrid computing ecosystem for practical quantum advantage. First, I will introduce my four research thrusts toward building this hybrid ecosystem: Classical Application Transformation, Adaptive Noise Mitigation, Scalable Error Correction and Efficient Resource Management. Second, from the Classical Application Transformation thrust, I will present "CAFQA: A classical simulation bootstrap for variational quantum algorithms", which enables accurate classical initialization for VQAs by searching efficiently through the classically simulable portion of the quantum space with Bayesian Optimization. CAFQA recovers as much as 99.99% of the accuracy lost in prior state-of-the-art classical initialization, with mean improvements of 56x. Third, from the Scalable Error Correction thrust, I will present "Clique: Better than worst-case decoding for quantum error correction", which proposes the Clique QEC decoder for cryogenic quantum systems. Clique is a lightweight cryo-decoder for decoding and correcting common trivial errors, so that only the rare complex errors are handled outside the cryo-refrigerator. Clique eliminates 90-99+% of the cryo-refrigerator I/O decoding bandwidth, while supporting more than a million physical qubits. Finally, I will conclude with an overview of other prior and ongoing work, along with my future research vision toward practical quantum advantage. Location: Core 301 Committee: |
| Start Date: 27 Feb 2023; Start Time: 09:00AM - 11:00AM Title: ROOTS: Object-Centric Representation and Rendering of 3D Scenes Bio: Speaker: Abstract: See above Location: Hill Center 350 (IDEAS Lounge) Committee: Professor Sungjin Ahn (Advisor) Professor Hao Wang Professor Karl Stratos Professor Xiong Fan |
| Start Date: 27 Feb 2023; Start Time: 10:30AM - 11:30AM Title: Convex Integer Optimization Bio: Haotian Jiang is a Postdoctoral Researcher at Microsoft Research, Redmond. In December 2022, he obtained his PhD from the Paul G. Allen School of Computer Science & Engineering at University of Washington under the supervision of Yin Tat Lee. He is broadly interested in theoretical computer science and applied mathematics. His primary area of expertise is the design and analysis of algorithms for continuous and discrete optimization problems. His work on optimization has been recognized by a Best Student Paper Award in SODA 2021. Speaker: Abstract: Designing and analyzing algorithms for optimization problems is a crucial but challenging task that arises in various fields such as business, science, and engineering. Despite the development of various successful optimization algorithms over the past sixty years, many state-of-the-art algorithms are theoretically far from optimal and ad hoc in nature. In this talk, I will present a unified toolbox for designing optimization algorithms through the general problem of Convex Integer Optimization, which captures many central problems and challenges in optimization today. Our toolbox has resulted in faster algorithms for fundamental tractable problems and better approximation algorithms for central NP-hard problems. Furthermore, it has revealed new connections between NP-hard and tractable problems which have been studied relatively independently for over half a century. Our work has created avenues for further investigations and applications in other fields of computer science and operations research, such as understanding the security of lattice-based cryptography and creating faster solvers for integer programming. Finally, I will conclude the talk with several future research directions and open problems at the frontier of optimization and related areas. Location: Core 301 Committee: |
| Start Date: 02 Mar 2023; Start Time: 08:30AM - 10:30AM Title: Matchings in Evolving Graphs Bio: Speaker: Abstract: See above Location: CoRE 431 Committee: Professor Aaron Bernstein (Chair) Professor Jie Gao Professor Sepehr Assadi Professor Sayan Bhattacharya (University of Warwick) |
| Start Date: 02 Mar 2023; Start Time: 10:30AM - 11:30AM Title: Data Structures for Fast Systems Bio: Alex Conway is a senior researcher at VMware. He received his PhD from Rutgers, where he was advised by Martín Farach-Colton. His work has primarily focused on randomized data structures and their use in storage systems, and covers the full research stack, from theory to systems to product. He is the co-creator and research lead of SplinterDB, an enterprise-grade key-value store deployed in VMware products. Speaker: Abstract: In this talk, I'll show how algorithms can be used to solve decades-old problems in systems design. I'll present an algorithmic approach to co-designing TLB hardware and the paging mechanism to increase TLB reach without the fragmentation issues incurred by huge pages. Along the way, I'll introduce a new hash-table design that overcomes existing tradeoffs, and achieves better performance than state-of-the-art hash tables both in theory and in practice. Key to these results are "tiny pointers," an algorithmic technique for compressing pointers. Location: Core 301 Committee: |
| Start Date: 03 Mar 2023; Start Time: 10:30AM - 11:30AM Title: Designing Exascale Distributed Systems Bio: Saurabh Kadekodi obtained his PhD in the Computer Science Department at Carnegie Mellon University (CMU) in 2020 as part of the Parallel Data Laboratory (PDL) under the guidance of Prof. Gregory Ganger and Prof. Rashmi Vinayak. After graduation Saurabh joined Google as a Visiting Faculty Researcher, and is currently a Research Scientist in the Storage Analytics team. Saurabh is broadly interested in designing distributed systems with special focus on the performance and reliability of storage systems. At Google, Saurabh is working towards implementing his PhD thesis on disk-adaptive redundancy and other exciting research ideas in some of the largest systems in the world. Speaker: Abstract: Fundamental physical limitations have slowed down hardware scaling, thus ending the “free”scaling benefits of processing power and storage capacity. At the same time, data is growing atan unprecedented rate. This data juggernaut is highly disruptive. It morphs benign assumptionsinto critical bottlenecks, and forces radical system (re-)designs. My work replaces designdecisions of distributed systems that are disrupted by scale with new, data-driven solutions thatare efficient, scalable, nimble, and robust. As an example, I will describe disk-adaptiveredundancy (DARE): a novel redesign of data reliability in exascale storage clusters driven byinsights gleaned from studying over 5.3 million disks from production environments of Google,NetApp and Backblaze. I will also describe three new DARE systems that reduce conservativeover-protection of data by up to 20% amounting to millions of dollars of cost savings along witha significant carbon footprint reduction, while always meeting desired data reliability targets.Additionally, I will briefly describe some past and current research efforts to improve theavailability and performance of local and distributed storage systems including new erasurecodes that reduce observed unavailability events at Google by up to 33%, a novel agingframework that can systematically age local file systems to look over 20 years old in less than 6hours, and an efficient packing and indexing layer in public cloud infrastructures that boosts thethroughput of accessing tiny objects by over 60x while simultaneously reducing the cost ofaccessing them by over 25000x. Finally, I will touch upon the open challenges in designingexascale distributed systems and highlight promising future directions. Location: CoRE 301 Committee: |
| Start Date: 03 Mar 2023; Start Time: 03:00PM - 04:00PM Title: Neuro-symbolic learning for bilevel planning Bio: Tom Silver is a fifth year PhD student at MIT EECS advised by Leslie Kaelbling and Josh Tenenbaum. His research is at the intersection of machine learning and planning with applications to robotics, and often uses techniques from task and motion planning, program synthesis, and reinforcement learning. Before graduate school, he was a researcher at Vicarious AI and received his B.A. from Harvard in computer science and mathematics in 2016. His work is supported by an NSF fellowship and an MIT presidential fellowship. Speaker: Abstract: Decision-making in robotics domains is complicated by continuous state and action spaces, long horizons, and sparse feedback. One way to address these challenges is to perform bilevel planning with abstractions, where a high-level search for abstract plans is used to guide planning in the original transition space. In this talk, I will give an overview of our recent efforts [1, 2, 3, 4] to design a bilevel planning system with state and action abstractions that are learned from data. I will also make the case for learning abstractions that are compatible with highly optimized PDDL planners, while arguing that PDDL planning should be only one component of a larger integrated planning system.[1] Learning symbolic operators for task and motion planning. Silver*, Chitnis*, Tenenbaum, Kaelbling, Lozano-Perez. IROS 2021.[2] Learning neuro-symbolic relational transition models for bilevel planning. Chitnis*, Silver*, Tenenbaum, Lozano-Perez, Kaelbling. IROS 2022.[3] Predicate invention for bilevel planning. Silver*, Chitnis*, Kumar, McClinton, Lozano-Perez, Kaelbling, Tenenbaum. AAAI 2023.[4] Learning neuro-symbolic skills for bilevel planning. Silver, Athalye, Tenenbaum, Lozano-Perez, Kaelbling. CoRL 2022. Location: 1 Spring Street, Room 403 Committee: |
| Start Date: 06 Mar 2023; Start Time: 10:30AM - 11:30AM Title: When fast algorithms meet modern society Bio: Omri Ben-Eliezer is an instructor (postdoc) in applied mathematics at MIT. He received his PhD in computer science from Tel Aviv University under the supervision of Prof. Noga Alon, and held postdoctoral positions at Weizmann Institute and Harvard University. His research blends aspects of algorithm design and data modeling, with specific interests including sublinear-time and streaming algorithms, large networks, robustness and privacy, and knowledge representation. For his work, Omri received several awards, including best paper awards at PODS 2020 and at CVPR 2020 Workshop on Text and Documents, the 2021 SIGMOD Research Highlight Award, and the first Blavatnik Prize for outstanding Israeli doctoral students in computer science. Speaker: Abstract: The rapidly growing societal impact of data-driven systems requires modern algorithms to process massive-scale complex data not just efficiently, but also responsibly, e.g., under privacy or robustness guarantees. In this talk I will discuss some of my recent research developing fast (e.g., sublinear-time or sublinear-space) and responsible algorithms for modern data analysis problems. I will focus on three representative lines of work: (i) the first systematic investigation of adversarial robustness in streaming algorithms, (ii) algorithm design in real-world social networks via new notions of core-periphery sparsification, and (iii) differentially private synthetic data generation in high dimensions via beyond-worst-case data modeling. Through these examples, I will demonstrate how the symbiosis between algorithm design and modeling of complex data often leads naturally to new structural insights and multidisciplinary connections Location: Core 301 Committee: |
| Start Date: 06 Mar 2023; Start Time: 02:00PM - 03:30PM Title: Efficient Datacenter Management Using Deep Reinforcement Learning Bio: Speaker: Abstract: See above Location: CoRE 301 Committee: Professor Thu Nguyen (Advisor) Professor He Zhu Professor Ulrich Kremer Professor David Pennock
|
| Start Date: 07 Mar 2023; Start Time: 10:30AM - 11:30AM Title: When Causal Inference Meets Graph Machine Learning: Unleashing the Potential of Mutual Benefit Bio: Jing Ma is a Ph.D. candidate in the Department of Computer Science at University of Virginia, under the supervision of Dr. Jundong Li and Dr. Aidong Zhang. She received her B.Eng. degree and M.Eng. degree at Shanghai Jiao Tong University with Outstanding Graduate Award. Her research interests broadly cover machine learning and data mining, especially include causal inference, graph mining, fairness, trustworthiness, and AI for social good. Her recent work focuses on bridging the gap between causality and machine learning. Her research papers have been published in top conferences and journals such as KDD, NeurIPS, IJCAI, WWW, AAAI, TKDE, WSDM, SIGIR, ECML-PKDD, AI Magazine, and IPSN. She has rich internship experience in companies and academic organizations such as Microsoft Research. She has won some important awards such as SIGKDD 2022 Best Paper Award and CAPWIC 2022 Best Poster Award. Speaker: Abstract: Recent years have witnessed rapid development in graph-based machine learning (ML) in various high-impact domains (e.g., healthcare, recommendation, and security), especially those powered by effective graph neural networks (GNNs). Currently, the mainstream graph ML methods are based on statistical learning, e.g., utilizing the statistical correlations between node features, graph structure, and labels for node classification. However, statistical learning has been widely criticized for only capturing the superficial relations between variables in the data system, and consequently, rendering the lack of trustworthiness in real-world applications. For example, ML models often make biased predictions toward underrepresented groups. Besides, these ML models often lack explanation for human. Therefore, it is crucial to understand the causality in the data system and the learning process. Causal inference is the discipline that investigates the causality inside a system, for example, to identify and estimate the causal effect of a certain treatment (e.g., wearing a face mask) on an important outcome (e.g., COVID-19 infection). Involving the concepts and philosophy of causal inference into ML methods is often considered as a significant component of human-level intelligence and can serve as the foundation of artificial intelligence (AI). However, most traditional causal inference studies rely on strong assumptions, and focus on independent and identically distributed (i.i.d.) data, while causal inference on graphs is faced with many barriers in effectiveness. Fortunately, the interplay between causal inference and graph ML has the potential to bring mutual benefit to each other. In this talk, we will present the challenges and our contributions for bridging the gap between causal inference and graph ML, mainly including two directions: 1) leveraging graph ML methods to facilitate causal inference in effectiveness; and 2) leveraging causality to facilitate graph ML models in model trustworthiness (e.g., model fairness and explanation). Location: Core 301 Committee: |
| Start Date: 16 Mar 2023; Start Time: 03:00PM - 05:00PM Title: Deep Learning with Limited Data Bio: Speaker: Abstract: See above. Location: CBIM 22 Committee: Professor Dimitris Metaxas (Chair) Professor HaoWang Professor Konstantinos Michmizos Professor Sharon Xiaolei Huang (Pennsylvania State University)
|
| Start Date: 20 Mar 2023; Start Time: 10:30AM - 12:00PM Title: Rescuing Data Center Processors Bio: Tanvir Ahmed Khan is a final-year Ph.D. candidate at the University of Michigan. His research brings together computer architecture, compilers, and operating systems to enable efficient data center processing. Consequently, his work has been adopted by Intel and ARM to improve the performance of data center applications. Bridging hardware and software, his research appears in venues like ISCA, MICRO, OSDI, PLDI, FAST, and EuroSys. His work has also been recognized with the MICRO 2022 Best Paper Award, DATE 2023 Best Paper Award Nomination, IEEE Micro Top Picks 2023 distinction, and multiple fellowships. Speaker: Abstract: Billions of people rely on web services powered by data centers, where critical applications run 24/7. Unfortunately, data center applications are extremely inefficient, wasting more than 60% of all processor cycles, and causing millions of dollars in operational expenses and energy costs. In this talk, I will present an overview of my vision to overcome this inefficiency using hardware/software co-design. In particular, I will focus on (1) systems interfaces using which software can reason about hardware inefficiencies; and (2) architectural abstractions using which software can avoid hardware inefficiencies. Finally, I will conclude by describing my ongoing and future research on democratizing hardware/software co-design to enable efficiency across the systems stack. Location: CoRE 301 Committee: |
| Start Date: 21 Mar 2023; Start Time: 10:30AM - 11:30AM Title: AI for Market and Policy Design: Integrating Data, Algorithms, and Economic Modeling Bio: Xintong Wang is a postdoctoral fellow at Harvard University, School of Engineering and Applied Sciences, working with David Parkes. Her research interests lie in understanding agent incentives and behaviors for the efficient design of multi-agent systems, using tools from AI and economics.
Xintong received her Ph.D. in Computer Science from the University of Michigan, advised by Michael Wellman. She has worked as a research intern at Microsoft Research and J.P. Morgan AI Research. She was selected as a Rising Star in EECS by UIUC in 2019 and a Rising Star in Data Science by the University of Chicago in 2022. Previously, Xintong received her B.S. with honors from Washington University in St. Louis in 2015. Speaker: Abstract: Today's markets have become increasingly algorithmic, with participants (i.e., agents) using algorithms to interact with each other at an unprecedented complexity, speed, and scale. Prominent examples of such algorithms include dynamic pricing algorithms, recommender systems, advertising technology, and high-frequency trading. Despite their effectiveness in achieving individual goals, the algorithmic nature poses challenges in designing economic systems that can align individual behavior to social objectives.This talk will highlight our work that tackles these challenges using tools from AI and economics, towards a vision of constructing efficient and healthy market-based, multi-agent systems. I will describe how we combine machine learning with economic modeling to understand strategic behaviors observed in real-world markets, analyze incentives behind such behaviors under game-theoretic considerations, and reason about how agents will behave differently in the face of new designs or environments. I will discuss the use of our method in two settings: (1) understanding and deterring manipulation practices in financial markets and (2) informing regulatory interventions that can incentivize a platform (e.g., Uber Eats) to act, in the forms of fee-setting and matching, to promote the efficiency, user diversity, and resilience of the overall economy. I will conclude by discussing future directions (e.g., model calibration, interpretability, scalability, and behavioral vs. rational assumptions) in using AI for the modeling and design of multi-agent systems. Location: Core 301 Committee: |
| Start Date: 22 Mar 2023; Start Time: 10:30AM - 11:30AM Title: Towards Inclusive and Equitable Language Technologies Bio: Malihe Alikhani is an Assistant Professor of computer science in the School of Computing and Information at the University of Pittsburgh. She earned her Ph.D. in computer science with a graduate certificate in cognitive science from Rutgers University in 2020 under the supervision of Prof. Matthew Stone. Her research interests center on using representations of communicative structure, machine learning, and cognitive science to design equitable and inclusive NLP systems for critical applications such as education, health, and social justice. Her work has received multiple best paper awards at ACL 2021, UAI2022, INLG2021, and UMAP2022 and has been supported by DARPA, NIH, Google, and Amazon. Speaker: Abstract: With the increasing deployment of language technologies to users, the need for a deeper understanding of the impact of natural language processing models on our society and user behaviors has grown. Designing culturally responsible, equitable, and inclusive language technologies that can benefit a diverse population is ever more important. Toward this goal, I present two directions: 1) Discourse-aware models for more inclusive social media moderation methods, and 2) Equitable machine learning frameworks for multimodal communication. Finally, I describe my research vision: building inclusive and collaborative communicative systems by leveraging the cognitive science of language use alongside formal machine learning methods. Location: CBIM 22 Committee: |
| Start Date: 23 Mar 2023; Start Time: 10:30AM - 12:00PM Title: Designing Formally Correct Intermittent Systems Bio: Milijana Surbatovich is a PhD Candidate in the Electrical and Computer Speaker: Abstract: "Extreme edge computing" is an emerging computing paradigm targeting application domains like medical wearables, disaster-monitoring tiny satellites, or smart infrastructure. This paradigm brings sophisticated sensing and data processing into an embedded device's deployment environment, enabling computing in environments that are too harsh, inaccessible, or dense to support frequent communication with a central server. Batteryless, energy harvesting devices (EHDs) are key to enabling extreme edge computing; instead of using batteries, which may be too costly or even impossible to replace, they can operate solely off energy collected from their environment. However, harvested energy is typically too weak to power a device continuously, causing frequent, arbitrary power failures that break traditional software and make correct programming difficult. Given the high assurance requirements of the envisioned application domains, EHDs must execute software without bugs that could render the device inoperable or leak sensitive information. While researchers have developed intermittent systems to support programming EHDs, they rely on informal, undefined correctness notions that preclude proving such necessary correctness and security properties.My research lays the foundation for designing formally correct intermittent systems that provide correctness guarantees. In this talk, I show how existing correctness notions are insufficient, leading to unaddressed bugs. I then present the first formal model of intermittent execution, along with correctness definitions for important memory consistency and timing properties. I use these definitions to design and implement both the language abstractions that programmers can use to specify their desired properties and the enforcement mechanisms that uphold them. Finally, I discuss my future research directions in intermittent system security and leveraging formal methods for full-stack correctness reasoning. Location: CoRE 301 Committee: |
| Start Date: 24 Mar 2023; Start Time: 10:30AM - 11:30AM Title: Towards understanding and defending against algorithmically curated misinformation Bio: Prerna Juneja is a social computing researcher at the University of Washington. She combines data-driven techniques, and human-centered design processes with large-scale real-world deployments to understand and then build defenses against problematic online phenomena. Prerna’s research has won multiple awards at human-computer interaction venues, such as ACM CHI, and ACM CSCW, and has been funded by the prestigious Social Data Research and Dissertation Fellowship. Her work has also received widespread press coverage by notable news channels, such as The Guardian, Al Jazeera, Sky News, and Seattle Times. Speaker: Abstract: Search engines and social media platforms are an indispensable part of our lives. Despite their increasing popularity, their search, ranking, and recommendation algorithms remain a black box to the users. The relevance of results produced by these search engines is driven by market factors and not by the quality of the content of those results. There is no guarantee that the information presented to people on online platforms is credible. To make matters worse, there are increasing concerns that online platforms amplify inaccurate information, making it easily accessible via search results and recommendations. My research takes a step towards understanding and designing defenses against algorithmically curated and amplified misinformation. In this talk, I will first present the results of a series of algorithmic audits I performed on online platforms to determine the extent to which algorithms contribute to the spread of misinformation and under which conditions they do so. Second, I'll present the design of an online system that aims to monitor an online platform for misinformation, developed in collaboration with several fact-checking organizations. Finally, I will discuss the opportunities in the space of algorithm auditing and the various ways in which we can redress the harm caused by the algorithms. Location: Core 301 Committee: |
| Start Date: 27 Mar 2023; Start Time: 10:30AM - 11:30AM Title: Distance-Estimation in Modern Graphs: Algorithms and Impossibility Bio: Nicole Wein is a Simons Postdoctoral Leader at DIMACS at Rutgers University. Previously, she obtained her Ph.D. from MIT advised by Virginia Vassilevska Williams. She is a theoretical computer scientist and her research interests include graph algorithms and lower bounds including in the areas of distance-estimation algorithms, dynamic algorithms, and fine-grained complexity. Speaker: Abstract: The size and complexity of today's graphs present challenges that necessitate the discovery of new algorithms. One central area of research in this endeavor is computing and estimating distances in graphs. In this talk I will discuss two fundamental families of distance problems in the context of modern graphs: Diameter/Radius/Eccentricities and Hopsets/Shortcut Sets. The best-known algorithm for computing the diameter (largest distance) of a graph is the naive algorithm of computing all-pairs shortest paths and returning the largest distance. Unfortunately, this can be prohibitively slow for massive graphs. Thus, it is important to understand how fast and how accurately the diameter of a graph can be approximated. I will present tight bounds for this problem via conditional lower bounds from fine-grained complexity. Secondly, for a number of settings relevant to modern graphs (e.g. parallel algorithms, streaming algorithms, dynamic algorithms), distance computation is more efficient when the input graph has low hop-diameter. Thus, a useful preprocessing step is to add a set of edges (a hopset) to the graph that reduces the hop-diameter of the graph, while preserving important distance information. I will present progress on upper and lower bounds for hopsets. Location: Core 301 Committee: |
| Start Date: 27 Mar 2023; Start Time: 02:00PM - 04:00PM Title: Bioinspired Neuromorphic Motion Control for Robots and Animated Characters Bio: Speaker: Abstract: See above Location: CoRE 301 Committee: Professor Mridul Aanjaneya (Co-Adviser) Professor Konstantinos Michmizos (Co-Adviser) Professor Kostas Bekris Professor Abdeslam Boularias Professor Tamar Shinar (UC Riverside)
|
| Start Date: 27 Mar 2023; Start Time: 04:00PM - 06:00PM Title: Diffusion Guided Image Generator Domain Adaptation Bio: Speaker: Abstract: See above Location: CBIM 22 Committee: Professor Ahmed Elgammal (Advisor) Professor Dimitris Metaxas Professor Hao Wang Professor Mario Szegedy
|
| Start Date: 28 Mar 2023; Start Time: 10:30AM - 11:30AM Title: Navigating the Challenges of Algorithmic Decision Making: Fair and Robust Automated Systems for Low-Resource Communities Bio: Arpita Biswas is currently a Research Associate at the Harvard T.H. Chan School of Public Health. Prior to this, she was a CRCS Postdoctoral Fellow at the John A. Paulson School of Engineering and Applied Sciences, Harvard University. She earned her Ph.D. degree from the Department of Computer Science and Automation (CSA), Indian Institute of Science (IISc). She has been awarded the Best Thesis Prize by the Indian Academy of Engineering (INAE) 2021, Best Thesis Award by the Department of CSA, IISc (2020-2021), a Google Ph.D. Fellowship (2016-2020), and a GHCI scholarship (2018). She has been recognized as a Rising Star in STOC 2021 and in the Trustworthy ML Seminar Series for her contribution to algorithms for fair decision-making. Her broad areas of interest include Algorithmic Game Theory, Optimization, and Machine Learning. She has worked on a wide range of problems, including fair resource allocation, health intervention planning, multi-agent learning, and robust sequential decision making. More details about her can be obtained at https://sites.google.com/view/arpitabiswas/. Speaker: Abstract: Automated decision-making systems play an increasingly important role in many societal decisions, such as health intervention planning, resource allocation, loan approvals, and criminal risk assessments. However, ensuring the responsible use of these systems is a challenging problem, especially for under-represented and low-resource communities. In this talk, I’ll present my research on fair and robust algorithms under resource limitations and other problem-specific constraints. The talk will cover two main themes: (1) Fair decision making in allocation and recommendation: Fairness is an important consideration in scenarios where a limited set of discrete resources is distributed among several agents, each having their own preferences. Two well-studied fairness notions in this context are envy-freeness up to one item (EF1) and maximin share (MMS). I have investigated the existence of these fairness notions under various constrained settings and developed algorithms that satisfy these fairness criteria. Further, I have used these solution concepts to quantify and promote fairness in two-sided markets (such as Netflix and Spotify) comprising customers on one side, and producers of goods/services on the other. Traditionally, recommendation services in these platforms have focused on maximizing customer satisfaction by tailoring the home-screen results according to the personalized preferences of users. However, our investigation reveals that such customer-centric recommendations may lead to unfair distribution of exposure among the producers, who may depend on such platforms to earn a living. I established that the two-sided fair recommendation problem can be reduced to the problem of constrained fair allocation of indivisible goods. I developed an algorithm, FairRec, that ensures maximin threshold guarantee in the exposure for a majority of the producers, and EF1 fairness for all the customers. Extensive evaluations over multiple real-world datasets show the effectiveness of FairRec in ensuring two-sided fairness while incurring only a marginal reduction in overall recommendation quality. (2) Robust sequential intervention planning: In many public health settings, it is important to provide interventions to ensure that patients adhere to health programs, such as taking medications and periodic health checks. This is extremely crucial among low-income communities who have limited access to preventive care information and healthcare facilities. In India, a non-profit called ARMMAN provides free automated voice messages to spread preventive care information among pregnant women. One of the key challenges is to ensure that the enrolled women continue listening to the voice messages throughout their pregnancy and after childbirth. Disengagements are detrimental to their health since they often have no other source for receiving timely healthcare information. While systematic interventions, such as scheduling in-person visits by healthcare workers, can help increase their listenership, interventions are often expensive and can be provided to only a small fraction of the enrolled women. I model this as a restless multi-armed bandit (RMAB) problem, where each beneficiary is assumed to transition from one state to another depending on the intervention provided to them. I developed a robust algorithm to tackle the uncertainty in transition dynamics and can potentially reduce the number of missed voice messages by 50%. I will conclude by delving into the best practices for responsible automated decision-making and discussing future research directions. I aim to showcase my overarching vision of fostering fairness, robustness, and scalability in the realm of automated decision-making through collaboration and continuous innovation. Location: Core 301 Committee: |
| Start Date: 29 Mar 2023; Start Time: 01:30PM - 03:00PM Title: Programmatic Reinforcement Learning Bio: Speaker: Abstract: See above Location: CoRE 301 Committee: Prof. He Zhu (advisor) Prof. Shiqing Ma Prof. Srinivas Narayana Prof. Qiong Zhang
|
| Start Date: 31 Mar 2023; Start Time: 10:30AM - 11:30AM Title: Responsible AI via Responsible Large Language Models Bio: Sharon is a 5th-year Ph.D. candidate at the University of California, Santa Barbara, where she is advised by Professor William Wang. Her research interests lie in natural language processing, with a focus on Responsible AI. Sharon’s research spans the subareas of fairness, trustworthiness, and safety, with publications in ACL, EMNLP, WWW, and LREC. She has spent summers interning at AWS, Meta, and Pinterest. Sharon is a 2022 EECS Rising Star and a current recipient of the Amazon Alexa AI Fellowship for Responsible AI. Speaker: Abstract: Large language models have advanced the state-of-the-art in natural language processing and achieved success in tasks such as summarization, question answering, and text classification. However, these models are trained on large-scale datasets, which may include harmful information. Studies have shown that as a result, the models exhibit social biases and generate misinformation after training. In this talk, I will discuss my work on analyzing and interpreting the risks of large language models across the areas of fairness, trustworthiness, and safety. I will first describe my research in the detection of dialect bias between African American English (AAE) vs. Standard American English (SAE). The second part will investigate the trustworthiness of models through the memorization and subsequent generation of conspiracy theories. The final part will discuss recent work in AI safety regarding text that may lead to physical harm. I will conclude my talk with discussions of future work in the area of Responsible AI. Location: Core 301 Committee: |
| Start Date: 03 Apr 2023; Start Time: 10:30AM - 11:30AM Title: Building Robust Systems Through System-Memory Co-Design Bio: Minesh Patel is a recent Ph.D. graduate from ETH Zurich. His thesis work focused on overcoming performance, reliability, and security challenges in the memory system. In particular, his dissertation identifies and addresses new challenges for system-level error detection and mitigation targeting memory chips with integrated error correcting codes (ECC). He also worked collaboratively on understanding and solving the RowHammer vulnerability, near-data processing, efficient virtual memory management, and new hardware security primitives. Throughout his graduate career, Minesh's contributions have been recognized with several honors, including DSN'19 and MICRO'20 Best Paper Awards, the William Carter Dissertation Award in Dependability, the ETH Doctoral Medal, and induction into the ISCA Hall of Fame. Minesh has also been actively involved as a teaching assistant and research mentor for undergraduate and graduate students, several of whose research projects resulted in successful publications. Minesh is passionate about addressing robustness and security design challenges for current and emerging computing systems (e.g., autonomous systems, quantum computers) and is committed to teaching and education. Speaker: Abstract: Main memory (DRAM) plays a central role in shaping the performance, reliability, and security of a wide range of modern computing systems. Unfortunately, worsening DRAM scaling challenges (e.g., increasing single-bit error rates, growing RowHammer vulnerability) are a significant threat to building robust systems. In this talk, I will discuss our recent efforts to understand and overcome the system-wide dependability consequences of DRAM on-die error-correcting codes (on-die ECC), a self-contained proprietary error-mitigation mechanism that is prevalent within modern DRAM chips and widely employed throughout computing systems today. Through a combination of real-chip experiments, statistical analyses, and simulation studies, I will: (i) show that on-die ECC obfuscates the statistical properties of main memory errors in a manner specific to the on-die ECC implementation used by a given chip and (ii) build a detailed understanding of how this obfuscation occurs, what its consequences are, and how those consequences can be overcome through system-memory co-design. Finally, I will discuss future research directions that explore practical cross-stack solutions for building robust next-generation computing systems. Location: CoRE 301 Committee: |
| Start Date: 04 Apr 2023; Start Time: 10:30AM - 11:30AM Title: Building Better Data-Intensive Systems Using Machine Learning Bio: Ibrahim Sabek is a postdoc at MIT and an NSF/CRA Computing Innovation Fellow. He is interested in Speaker: Abstract: Database systems have traditionally relied on handcrafted approaches and rules to store large-scale data and process user queries over them. These well-tuned approaches and rules work well for the general-purpose case, but are seldom optimal for any actual application because they are not tailored for the specific application properties (e.g., user workload patterns). One possible solution is to build a specialized system from scratch, tailored to each application's needs. Although such a specialized system is able to get orders-of-magnitude better performance, building it is time-consuming and requires a substantial manual effort. This pushes the need for automated solutions that abstract system-building complexities while getting as close as possible to the performance of specialized systems.In this talk, I will show how we leverage machine learning to instance-optimize the performance of query scheduling and execution operations in database systems. In particular, I will show how deep reinforcement learning can fully replace a traditional query scheduler. I will also show that—in certain situations—even simpler learned models, such as piece-wise linear models approximating the cumulative distribution function (CDF) of data, can help improve the performance of fundamental data structures and execution operations, such as hash tables and in-memory join algorithms. Location: CoRE 301 Committee: |
| Start Date: 06 Apr 2023; Start Time: 03:30PM - 05:00PM Title: Enhancing Language Models with Logical Reasoning and Automatic Error Analysis Bio: Speaker: Abstract: See above Location: CoRE 305 Committee: Professor Yongfeng Zhang (Advisor) Professor He Zhu Professor Hao Wang Professor Peng Zhang
|
| Start Date: 14 Apr 2023; Start Time: 10:00AM - 12:00PM Title: Integrate Logical Reasoning and Machine Learning for Decision Making Bio: Speaker: Abstract: See above Location: CoRE 305 Committee: Professor Yongfeng Zhang (Advisor) Professor Hao Wang Professor Dong Deng Professor Sudarsan Kannan
|
| Start Date: 19 Apr 2023; Start Time: 03:00PM - 04:30PM Title: CrossPrefetch: Accelerating I/O Prefetching for Modern Storage Bio: Speaker: Abstract: See above Location: CoRE 301 Committee: Professor Sudarsan Kannan (Advisor) Professor Srinivas Narayana Professor Badri Nath Professor Karthick CS Professor Parashar Manish (University of Utah)
|
Subscribe to RSS FeedNews & Events
Upcoming Events
| 27 Mar 2023; - 10:30AM - 11:30AM Distance-Estimation in Modern Graphs: Algorithms and Impossibility |
| 27 Mar 2023; - 02:00PM - 04:00PM Bioinspired Neuromorphic Motion Control for Robots and Animated Characters |
| 27 Mar 2023; - 04:00PM - 06:00PM Diffusion Guided Image Generator Domain Adaptation |
Department of Computer Science