August 2022 - talks - lists.cs.princeton.edu

CS Colloquium Series: week of April 11-15
by Emily C. Lawrence 19 Sep '22

19 Sep '22

Here is the full list of CS Colloquium talks for next week. All talks will be recorded. ~~~~~ CS Colloquium Speaker Speaker: Parastoo Abtahi, Stanford University Date: Monday, April 11, 2022 Time: 12:30pm EST Location: CS 105 Hosts: Andrés Monroy-Hernández & Adam Finkelstein Event page: [ https://www.cs.princeton.edu/events/26186 | https://www.cs.princeton.edu/events/26186 ] This talk will be live-streamed at [ https://mediacentrallive.princeton.edu/ | https://mediacentrallive.princeton.edu/ ] Title: From Haptic Illusions in Virtual Reality to Beyond-Real Interactions Abstract: Advances in audiovisual rendering have led to the commercialization of virtual reality (VR) hardware; however, haptic technology has not kept up with these advances. While haptic devices aim to bridge this gap by simulating the sensation of touch, there are many hardware limitations that make realistic touch interactions in VR challenging. In my research, I explore how by understanding human perception, we can design VR interactions that not only overcome the current limitations of VR hardware, but also extend our abilities beyond what is possible in the real world. In this talk, I will present my work on redirection illusions that leverage the limits of human perception to improve the perceived performance of encountered-type haptic devices, such as improving the position accuracy of drones, the speed of tabletop robots, and the resolution of shape displays when used for haptics in VR. I will then present a framework I have developed through the lens of sensorimotor control theory to argue for the exploration and evaluation of VR interactions that go beyond mimicking reality. Bio: Parastoo Abtahi is a final year computer science PhD candidate and a Gerald J. Lieberman fellow at Stanford University, where she is co-advised by Prof. James Landay and Prof. Sean Follmer. Her research area is human-computer interaction (HCI) and she works broadly on virtual reality interactions and spatial computing. Her research has been published at top HCI venues, such as the ACM Conference on Human Factors in Computing Systems (CHI) and the ACM User Interface Software and Technology Symposium (UIST), and has received two honorable mention paper awards. Her work has been supported by the Stanford Institute for Human-Centered Artificial Intelligence, the Hasso Plattner Design Thinking Research Program, and the VMware fellowship. Prior to Stanford, Parastoo received her bachelor’s degree in Electrical and Computer Engineering from the University of Toronto as part of the Engineering Science program. ~~~~~ CS Colloquium Speaker Speaker: Ferran Alet, Massachusetts Institute of Technology Date: Tuesday, April 12, 2022 Time: 12:30pm EST Location: CS 105 Host: Olga Russakovsky Event page: https://www.cs.princeton.edu/events/26187 This talk will be live-streamed at https://mediacentrallive.princeton.edu/ Title: A flexible framework for machine learning Abstract: In this last decade, we have seen a lot of progress in AI and machine learning using a single recipe: given a task, we train a single neural network to map inputs to outputs. In this talk, I will show that this one-neural-network-per-task framework can be extended to improve generalization. First, I will describe modular meta-learning, which achieves language-like generalization by training a set of composable neural modules. By having multiple neural networks per task, and multiple tasks per neural network, we are able to reuse information and achieve bigger data and computational efficiency. In the second part of my talk, I will describe tailoring, a general way of encoding inductive biases in neural networks by optimizing unsupervised objectives inside the prediction function, essentially having one neural network per input. Finally, I will describe my vision for creating a flexible ML framework that will enable training reinforcement learning policies within minutes rather than days, solving complex search and discovery problems, and improving our understanding of generalization in deep learning. Bio: Ferran Alet is a PhD candidate at MIT CSAIL advised by Leslie Kaelbling, Tomas Lozano-Perez, and Joshua Tenenbaum. His research is on machine learning and leverages techniques from meta-learning, learning to search, program synthesis, and insights from mathematics and the physical sciences. During his PhD, he created the MIT Embodied Intelligence Seminar, mentored 17 students, and won the MIT Outstanding Mentor award 2021. Ferran studied mathematics and physics in Barcelona thanks to CFIS, a program for doing two degrees, where he was the valedictorian of his promotion. In college, he participated in the ACM-ICPC programming contest, being the most decorated in the history of his regional phase (South Western Europe). In grad school, he earned a “La Caixa” fellowship and was responsible for the high-level planner of the MIT-Princeton team for the Amazon Robotics Challenge, which won the stowing task in 2017. You can find more information and papers at www.alet-et.al ~~~~~ CS Colloquium Speaker Speaker: Benedikt Bünz, Stanford University Date: Thursday, April 14, 2022 Time: 12:30pm EST Location: CS 105 Hosts: Jonathan Mayer & Arvind Narayanan Event page: https://www.cs.princeton.edu/events/26188 This talk will be live-streamed at https://mediacentrallive.princeton.edu/ Title: Improving the privacy, scalability, and ecological impact of blockchains Abstract: Blockchains are an exciting area of research that touches on many areas of Computer Science and beyond. This technology has the potential to enable a fast, cheap, and private financial system based on distributed consensus and cryptography, instead of trusted parties. Despite this potential, the reality still shows severe limitations of blockchains: (i) transactions can cost hundreds of dollars and take minutes to confirm, (ii) some blockchains offer little privacy, and (iii) proof-of-work consensus consumes too much energy. In this talk, I will discuss powerful techniques that follow a prover paradigm and can mitigate these limitations. The first technique, called Bulletproofs, is a general-purpose zero-knowledge proof system that is specifically designed to enable confidential blockchain transactions. Bulletproofs requires minimal trust assumptions and gives the shortest zero-knowledge proofs without trusted setup. The system is widely deployed and powers tens of thousands of private blockchain transactions per day. The second technique, called inner pairing products, is a way to aggregate many zero-knowledge proofs into a single short proof. This can significantly reduce on-chain data, leading to a significant increase in transactions per second that the chain can process. The third technique is a new concept called a verifiable delay function (VDF) that is vital for permission-less and eco-friendly consensus. VDFs are already deployed in Filecoin and Chia, and are planned for Ethereum 2.0, the upcoming upgrade to Ethereum. Bio: Benedikt Bünz is a PhD candidate at Stanford University, a member of Dan Boneh’s applied cryptography lab, and a recipient of the Microsoft Research Fellowship at the Simons Institute. His work on the science of Blockchains uses tools from applied cryptography, distributed systems, and algorithmic game theory. His research focuses on building new proof protocols for improving the privacy, scalability, and ecological impact of blockchains. Several of his research results have had a significant industry impact. His work on Bulletproofs, secures tens of thousands of private transactions on Blockchains like Monero or Signal’s Mobilecoin. His seminal work on Verifiable Delay Functions (VDFs) sparked the VDF Alliance, a multi-million dollar initiative composed of academic, non-profit, and corporate collaborators.

1 11

Reminder: Princeton AI Club Kick-Off Event with Special Guest Speaker: Prof. Yann LeCun (Meta-FAIR & New York University)
by Emily C. Lawrence 30 Aug '22

30 Aug '22

Reminder: Happening this week! The Princeton AI Club is thrilled to announce its kick-off event for the coming academic year, and we have a special guest speaker for you – [ http://yann.lecun.com/ | Prof. Yann LeCun ] from Meta-FAIR & New York University! When? September 1 st at 11:30 am Eastern Time (ET) ! Where? Computer Science Building, Room 105, Princeton University, and virtually. Sign up [ https://forms.gle/BonERMY1GNbS1vaM6 | here ] ! An in-person brunch and social event will start at 11:00 am [ https://calendar.google.com/event?action=TEMPLATE&tmeid=MW04ZDliYzc2NWE2aDM… | add brunch event to calendar ] ! Hurry up, in person attendance has limited availability. Speaker: Prof. Yann LeCun (Meta-FAIR & New York University) Title: From Machine Learning to Autonomous Intelligence Date: September 1 st at 11:30 am Eastern Time (ET). Our brunch and social event will start at 11:00 am. Abstract: How could machines learn as efficiently as humans and animals? How could machines learn to reason and plan? How could machines learn representations of percepts and action plans at multiple levels of abstraction, enabling them to reason, predict, and plan at multiple time horizons? I will propose a possible path towards autonomous intelligent agents, based on a new modular cognitive architecture and a somewhat new self-supervised training paradigm. The centerpiece of the proposed architecture is a configurable predictive world model that allows the agent to plan. Behavior and learning are driven by a set of differentiable intrinsic cost functions. The world model uses a new type of energy-based model architecture called H-JEPA (Hierarchical Joint Embedding Predictive Architecture). H-JEPA learns hierarchical abstract representations of the world that are simultaneously maximally informative and maximally predictable. The corresponding working paper is available here: [ https://openreview.net/forum?id=BZ5a1r-kVsf | https://openreview.net/forum?id=BZ5a1r-kVsf ] Bio: Yann LeCun is VP & Chief AI Scientist at Meta and Silver Professor at NYU affiliated with the Courant Institute of Mathematical Sciences & the Center for Data Science. He was the founding Director of FAIR and of the NYU Center for Data Science. He received an Engineering Diploma from ESIEE (Paris) and a PhD from Sorbonne Université. After a postdoc in Toronto he joined AT&T Bell Labs in 1988, and AT&T Labs in 1996 as Head of Image Processing Research. He joined NYU as a professor in 2003 and Meta/Facebook in 2013. His interests include AI machine learning, computer perception, robotics and computational neuroscience. He is the recipient of the 2018 ACM Turing Award (with Geoffrey Hinton and Yoshua Bengio) for "conceptual and engineering breakthroughs that have made deep neural networks a critical component of computing", a member of the National Academy of Sciences, the National Academy of Engineering, the French Académie des Sciences.

1 0

Princeton AI Club Kick-Off Event with Special Guest Speaker: Prof. Yann LeCun (Meta-FAIR & New York University)
by Emily C. Lawrence 26 Aug '22

26 Aug '22

The Princeton AI Club is thrilled to announce its kick-off event for the coming academic year, and we have a special guest speaker for you – [ http://yann.lecun.com/ | Prof. Yann LeCun ] from Meta-FAIR & New York University! When? September 1 st at 11:30 am Eastern Time (ET) ! Where? Computer Science Building, Room 105, Princeton University, and virtually. Sign up [ https://forms.gle/BonERMY1GNbS1vaM6 | here ] ! An in-person brunch and social event will start at 11:00 am [ https://calendar.google.com/event?action=TEMPLATE&tmeid=MW04ZDliYzc2NWE2aDM… | add brunch event to calendar ] ! Hurry up, in person attendance has limited availability. Speaker: Prof. Yann LeCun (Meta-FAIR & New York University) Title: From Machine Learning to Autonomous Intelligence Date: September 1 st at 11:30 am Eastern Time (ET). Our brunch and social event will start at 11:00 am. Abstract: How could machines learn as efficiently as humans and animals? How could machines learn to reason and plan? How could machines learn representations of percepts and action plans at multiple levels of abstraction, enabling them to reason, predict, and plan at multiple time horizons? I will propose a possible path towards autonomous intelligent agents, based on a new modular cognitive architecture and a somewhat new self-supervised training paradigm. The centerpiece of the proposed architecture is a configurable predictive world model that allows the agent to plan. Behavior and learning are driven by a set of differentiable intrinsic cost functions. The world model uses a new type of energy-based model architecture called H-JEPA (Hierarchical Joint Embedding Predictive Architecture). H-JEPA learns hierarchical abstract representations of the world that are simultaneously maximally informative and maximally predictable. The corresponding working paper is available here: [ https://openreview.net/forum?id=BZ5a1r-kVsf | https://openreview.net/forum?id=BZ5a1r-kVsf ] Bio: Yann LeCun is VP & Chief AI Scientist at Meta and Silver Professor at NYU affiliated with the Courant Institute of Mathematical Sciences & the Center for Data Science. He was the founding Director of FAIR and of the NYU Center for Data Science. He received an Engineering Diploma from ESIEE (Paris) and a PhD from Sorbonne Université. After a postdoc in Toronto he joined AT&T Bell Labs in 1988, and AT&T Labs in 1996 as Head of Image Processing Research. He joined NYU as a professor in 2003 and Meta/Facebook in 2013. His interests include AI machine learning, computer perception, robotics and computational neuroscience. He is the recipient of the 2018 ACM Turing Award (with Geoffrey Hinton and Yoshua Bengio) for "conceptual and engineering breakthroughs that have made deep neural networks a critical component of computing", a member of the National Academy of Sciences, the National Academy of Engineering, the French Académie des Sciences.

1 0

Corey Sinnamon will present his FPO "Analysis of self-adjusting heaps" on Tuesday, 8/30/2022 at 3PM via Zoom.
by Gradinfo 24 Aug '22

24 Aug '22

Corey Sinnamon will present his FPO " Analysis of self-adjusting heaps" on Tuesday, 8/30/2022 at 3PM via Zoom. Zoom Link: [ https://princeton.zoom.us/my/sinnamon | https://princeton.zoom.us/my/sinnamon ] A copy of his thesis is available upon request. Please email gradinfo(a)cs.princeton.edu if you would like a copy of the thesis. Everyone is invited to attend his talk. The members of Corey's committee are as follows: Examiners: Robert Tarjan (Advisor), Bernard Chazelle, and Maria Chudnovsky. Readers: Huacheng Yu and László Kozma (Freie Universität Berlin). Abstract: We study the amortized efficiency of self-adjusting heap data structures. Heaps (also called priority queues) are fundamental data structures, and a great deal of work has been devoted to inventing efficient heaps and analyzing them. We prove new amortized bounds for pairing heaps, multipass pairing heaps, slim heaps, and smooth heaps. The pairing heap is a classic self-adjusting heap, known to be simple, robust, and efficient in practice. Despite decades of study, its exact efficiency is still unresolved. We prove new amortized bounds on the time per operation of the pairing heap: O(log n) amortized time for delete-min and delete; O( √ log log N2 √ 2 log2 log2 N ) amortized time for insert, meld, and decrease-key; and O(1) time for make-heap and find-min. Here n is the current number of items in the heap, and N is an upper bound on n across all operations. The multipass pairing heap is a well-known variant of the pairing heap. We prove it satisfies amortized time bounds of O(log n) for delete-min and delete; O(log log n·log log log n) for decreasekey; and O(1) for insert, meld, make-heap, and find-min. This is the first analysis in which both insert and delete-min take O(log n) amortized time. It matches the known lower bounds for all operations except decrease-key, and the bound for decrease-key comes close to the lower bound. We prove the same bounds for the lazy version of multipass pairing heaps, in which insert, meld, and decrease-key operations are done lazily. Again, these bounds are tight for all operations except decrease-key. The smooth heap and the closely related slim heap are much newer self-adjusting heaps. For these two heaps, we give a tight analysis of the amortized time per operation: O(log n) for deletemin and delete; O(log log n) for decrease-key; and O(1) for make-heap, find-min, insert, and meld. These bounds are tight not only for smooth and slim heaps, but for any heap in Iacono and Ozkan’s ¨ pure heap model, intended to capture all self-adjusting heaps. Slim and smooth heaps are the first known data structures to match Iacono and Ozkan’s lower bounds and to satisfy the constraints of ¨ their model. We prove tight bounds for both eager and lazy versions of slim and smooth heaps.

1 0

Matthew Myers will present his FPO "Inferring Intra-tumor Heterogeneity from DNA Sequencing Data" on Monday, August 29, 2022 at 11:00 AM in CS 301 and Zoom.
by Louis W. Riehl 22 Aug '22

22 Aug '22

Matthew Myers will present his FPO "Inferring Intra-tumor Heterogeneity from DNA Sequencing Data" on Monday, August 29, 2022 at 11:00 AM in CS 301 and Zoom. Location: Zoom link: [ https://princeton.zoom.us/j/94547797233?pwd=TmFsOXJUWllQSWVseUQ0RnJIeDh1QT09 | https://princeton.zoom.us/j/94547797233?pwd=TmFsOXJUWllQSWVseUQ0RnJIeDh1QT09 ] The members of Matthew’s committee are as follows: Examiners: Ben Raphael (Adviser), Yuri Pritykin, Mona Singh Readers: Olga Troyanskaya, Quaid Morris (Memorial Sloan Kettering Cancer Center) A copy of his thesis is available upon request. Please email gradinfo(a)cs.princeton.edu if you would like a copy of the thesis. Everyone is invited to attend his talk. Abstract follows below: Cancer is a disease characterized by somatic mutations which accumulate over time and in response to evolutionary pressures. As a result, tumors are composed of multiple distinct clones each characterized by a different set of mutations. This intra-tumor heterogeneity is closely related to negative outcomes such as treatment resistance, relapse, and metastasis. The somatic mutations that define tumor clones can range greatly in size, from single-nucleotide variants (SNVs) which change only a single base position to copy-number aberrations (CNAs) which can affect thousands of bases up to the whole genome. Researchers use DNA sequencing data to measure these somatic mutations and study intratumor heterogeneity. However, the vast majority of cancer sequencing uses DNA from bulk tumor samples which are mixtures of millions of cells. Thus, the resulting data is a combination of DNA sequences across all tumor and normal cells. This presents challenges for analysis, as it is not immediately apparent from the sequencing reads which somatic mutations characterize individual clones. Recently, single-cell sequencing technologies enable researchers to measure DNA sequencing reads from individual cells. However, these technologies have higher rates of sequencing errors and limited sequencing coverage, so sophisticated algorithms are still needed to recover the tumor clones and their mutations. In this dissertation, we present three computational methods for inferring tumor clones and their constituent mutations from either bulk or single-cell DNA sequencing data. The first method, CALDER, infers tumor clones and their evolutionary relationships using SNVs from longitudinal bulk DNA sequencing samples. CALDER uses the longitudinal ordering to apply constraints on the clones present in each sample. The second method, SBMClone, infers tumor clones using SNVs from ultra-low coverage single-cell DNA sequencing data using the stochastic block model, a well-studied tool from statistical physics and network science. The third method, HATCHet2, infers clones and their allele-specific CNAs from one or more bulk DNA sequencing samples. HATCHet2 improves upon the state of-the-art with several methodological innovations, including variable-width binning, locality-aware clustering, and a novel statistic for quantifying allelic imbalance which enables the identification of mirrored subclonal CNAs, in which different alleles are amplified in different clones. Louis Riehl Graduate Administrator Computer Science Department, CS213 Princeton University (609) 258-8014

1 0

Rachit Dubey will present his Pre-FPO "The successes and failures of intrinsic drives" on Wednesday, August 24, 2022 at 12:00 PM on Zoom.
by Louis W. Riehl 19 Aug '22

19 Aug '22

Rachit Dubey will present his Pre-FPO "The successes and failures of intrinsic drives" on Wednesday, August 24, 2022 at 12:00 PM on Zoom. Zoom link: [ https://princeton.zoom.us/j/95974138201 | https://princeton.zoom.us/j/95974138201 ] Committee Members: Thomas L. Griffiths (Examiner, advisor), Jonathan D. Cohen (Examiner), Karthik Narasimhan(examiner), Tania Lombrozo (reader), Ryan Adams (reader). All are welcome to attend. Title: The successes and failures of intrinsic drives Abstract: Even in the absence of external rewards like money or food, we have internal motives that drive and propel us to act. These intrinsic drives have led us to some of our greatest successes – they have motivated us to learn new things, explore new domains, and make some great discoveries. Unfortunately, left unchecked, these drives can be extremely harmful. This thesis aims to provide a deeper understanding of the successes and failures of human intrinsic drives by employing computational modeling and large-scale behavioral experiments. In the first half of the talk, we examine the drives that inspire us to pursue science, which is a remarkable success case of our drives. We first present a rational model of curiosity that explains a wide range of findings about human curiosity. Based on this model, we then present the results of some behavioral interventions that have the potential to pique people’s curiosity about everyday scientific topics. We then present a computational model of Aha! moments and provide an explanation for why Aha! moments feel so rewarding. In the second part of the talk, we examine a glaring failure case of our drives -- the problem of overconsumption. We first present a brief overview of the very urgent and pressing issue of overconsumption and put forth the argument that combating this problem requires a deeper understanding of our internal drives and reward functions. We then use the idea of reward design and show that the drives which cause overconsumption might paradoxically play an important role in promoting adaptive behavior. Louis Riehl Graduate Administrator Computer Science Department, CS213 Princeton University (609) 258-8014

1 0

Nikunj Saunshi will present his FPO "Towards Understanding Self-Supervised Representation Learning" on Friday, August 26, 2022 at 11:00 AM in CS 402 and on Zoom.
by Louis W. Riehl 19 Aug '22

19 Aug '22

Nikunj Saunshi will present his FPO " Towards Understanding Self-Supervised Representation Learning " on Friday, August 26, 2022 at 11:00 AM in CS 402 and on Zoom . Location: Zoom link: [ https://princeton.zoom.us/j/93697691837 | https://princeton.zoom.us/j/93697691837 ] The members of Nikunj ’s committee are as follows: Examiners: Sanjeev Arora (Adviser), Danqi Chen , Chi Jin Readers: Elad Hazan , Jason Lee A copy of his thesis is available upon request. Please email [ mailto:gradinfo@cs.princeton.edu | gradinfo(a)cs.princeton.edu ] if you would like a copy of the thesis. Everyone is invited to attend his talk. Abstract follows below: While supervised learning sparked the deep learning boom, it has some critical shortcomings: (1) it requires an abundance of expensive labeled data, and (2) it solves tasks from scratch rather than the human-like approach of leveraging knowledge and skills acquired from prior experiences. Pre-training has emerged as an alternative and effective paradigm, to overcome these shortcomings, whereby a model is first trained using easily acquirable data, and later used to solve downstream tasks of interest with much fewer labeled data than supervised learning. Pre-training using unlabeled data, a.k.a. self-supervised learning, has been especially revolutionary, with successes in diverse domains: text, vision, speech, etc. This raises an interesting and challenging question: why should pre-training on unlabeled data help with seemingly unrelated downstream tasks? In this thesis we present works that initiate and build a theoretical framework to study why self-supervised learning is beneficial for downs tream tasks. The framework is applied to methods like contrastive learning, auto-regressive language modeling and self-prediction based methods. Central to the framework is the idea that pre-training helps learn low-dimensional representations of data, that subsequently help solve downstream tasks of interest with linear classifiers, requiring fewer labeled data. A common theme is to formalize what are desirable properties of the unlabeled data distribution that is used to construct the self-supervised learning task. Under appropriate formalizations, it can be shown that approximately minimizing the right pre-training objectives can extract the downstream signal that is implicitly encoded in the unlabeled data distribution. Finally it is shown that this signal can be decoded from the learned representations using linear classifiers, thus providing a formalization for transference of “skills and knowledge” across tasks. Louis Riehl Graduate Administrator Computer Science Department, CS213 Princeton University (609) 258-8014

1 0

TODAY 8/15 at 10AM - ECE Seminar Saar Tochner, Hebrew University of Jerusalem
by Gradinfo 15 Aug '22

15 Aug '22

ECE SEMINAR Speaker: Saar Tochner, Hebrew University of Jerusalem Title: Twilight: A Differentially Private Payment Channel Network Day: August 15, 2022 Time: 10:00 am Location: B205 Engineering Quadrangle Hosts: Maria Apostolaki & Prateek Mittal Abstract: Payment channel networks (PCNs) provide a faster and cheaper alternative to transactions recorded on the blockchain. Clients can trustlessly establish payment channels with relays by locking coins and then sending signed payments that shift coin balances over the network's channels. Although payments are never published, anyone can track a client's payment by monitoring changes in coin balances over the network's channels. We present Twilight, the first PCN that provides a rigorous differential privacy guarantee to its users. Relays in Twilight run a noisy payment processing mechanism that hides the payments they carry. This mechanism increases the relay's cost, so Twilight combats selfish relays that wish to avoid it using a trusted execution environment (TEE) that ensures they follow its protocol. The TEE does not store the channel's state, which minimizes the trusted computing base. Crucially, Twilight ensures that even if a relay breaks the TEE's security, it cannot break the integrity of the PCN. We analyze Twilight in terms of privacy and cost and study the trade-off between them. We implement Twilight using Intel's SGX framework and evaluate its performance using relays deployed on two continents. We show that a route consisting of 4 relays handles 820 payments/sec. Bio: I'm a Ph.D. candidate advised by Aviv Zohar at the Hebrew University of Jerusalem. My research focuses on designing and understanding privacy and security aspects of systems, mainly in the field of cryptocurrencies. Website: http://saart.info/ This seminar is supported by ECE Korhammer Lecture Series Funds

1 0

Hei Law will present his FPO "Learning to Detect Objects by Grouping" on Tuesday, August 16, 2022 at at 10:00 AM in Friend 202 and Zoom.
by Louis W. Riehl 09 Aug '22

09 Aug '22

Hei Law will present his FPO "Learning to Detect Objects by Grouping" on Tuesday, August 16, 2022 at at 10:00 AM in Friend 202 and Zoom. Location: Zoom link: [ https://princeton.zoom.us/my/heilaw?pwd=a3U4ZmZBZXdTbWhqOFBHbmN1c2pEUT09 | https://princeton.zoom.us/my/heilaw?pwd=a3U4ZmZBZXdTbWhqOFBHbmN1c2pEUT09 ] The members of Hei’s committee are as follows: Examiners: Jia Deng (Adviser), Adam Finkelstein, Felix Heide Readers: Szymon Rusinkiewicz, Olga Russakovsky A copy of his thesis is available upon request. Please email gradinfo(a)cs.princeton.edu if you would like a copy of the thesis. Everyone is invited to attend his talk. Abstract follows below: Extracting high level semantic information from visual input is an ability that human rely on to perform daily tasks. This often requires identifying and locating relevant objects before any higher level information is inferred. This step is known as object detection which is a fundamental task in computer vision. It has numerous real world applications and serves as an upstream task for many computer vision tasks. This dissertation makes three contributions to object detection. First we propose CornerNet, a new approach to object detection. CornerNet reformulates object detection as detecting and grouping pairs of keypoints. More specifically, CornerNet detects corners of the bounding boxes and predicts similar embedding vectors for corners from the same objects. CornerNet also introduces corner pooling, a new pooling layer that helps localize corners. Experiments on COCO show that CornerNet achieves an AP of 42.2%, outperforming all one-stage detectors. Second we propose CornerNet-Lite, a collection of two efficient variants, CornerNet- Saccade and CornerNet-Squeeze, of CornerNet. CornerNet-Lite explores two orthog- onal directions: processing fewer pixels and reducing the processing cost of each pixel. Inspired by saccade in human vision system, CornerNet-Saccade processes fewer pix- els by estimating object locations on a downsampled image and processes a subet of regions in high resolution. It is 6x faster than CornerNet and achieves a better AP. CornerNet-Squeeze reduces the cost by introducing a new compact hourglass backbone network. It is faster and more accurate than YOLOv3. Third we propose “Synthetic Opitmized Layout with Instance Detection (SOLID)”, a new pretraining approach for object detection. SOLID consists of two main compo- nents. The first component generates synthetic images from a collection of unlabelled 3D models with optimized scene arrangement. The second component is an instance detection task where given a query image depicting a 3D model, a detector is trained to locate the instances of the same object in a target image. Experiments show that synthetic data can be effective for pretraining an object detector. >From grouping corners in CornerNet and CornerNet-Lite to grouping instances in SOLID, this dissertation presents a new approach to object detection – learning to detect objects by grouping. Louis Riehl Graduate Administrator Computer Science Department, CS213 Princeton University (609) 258-8014

1 0

Zhiyuan Li will present his FPO "Bridging Theory and Practice in Deep Learning: Optimization and Generalization" on Friday, August 12, 2022 at 1:30 PM on Zoom.
by Louis W. Riehl 08 Aug '22

08 Aug '22

Zhiyuan Li will present his FPO "Bridging Theory and Practice in Deep Learning: Optimization and Generalization" on Friday, August 12, 2022 at 1:30 PM on Zoom. Location: Zoom link: [ https://princeton.zoom.us/j/91892691202?pwd=aEVZYTA4cEZyc0ZTQWoxcDFWZHVWZz09 | https://princeton.zoom.us/j/91892691202?pwd=aEVZYTA4cEZyc0ZTQWoxcDFWZHVWZz09 ] The members of Zhiyuan’s committee are as follows: Examiners: Sanjeev Arora (Adviser), Chi Jin, Danqi Chen Readers: Elad Hazan, Jason Lee A copy of his thesis is available upon request. Please email gradinfo(a)cs.princeton.edu if you would like a copy of the thesis. Everyone is invited to attend his talk. Abstract follows below: Deep learning has been hugely successful for several important applications in the past decade, yet mathematical understanding has lagged behind its breathtaking empirical success. Classic machine learning theory is insufficient to explain various new phenomena in deep learning and to provide guidance on algorithmic choices, largely due to an oversimplified black box view that ignores the interaction between the model and the optimization algorithm. This dissertation presents a collection of theoretical results that take the interplay between the model and the optimization algorithm into account and aims to bridge the gaps between theory and practice in deep learning for both generalization and optimization. For optimization, we first illustrate the mismatches between traditional optimization theory and deep networks with normalization layers by presenting an exponentially increasing learning rate schedule that works well empirically. We explain this surprise by establishing its equivalence to SGD with Weight Decay and proving that their convergence rates are fast and insensitive to initialization scale. Based on this, we design a variant of BERT named SIBERT, which is trainable by SGD and thus more memory-efficient than adaptive algorithms like ADAM. Finally we present the first provable yet general setting where gradient descent decreases loss in a non-monotone way, as observed empirically. For generalization, we study the implicit bias of optimization algorithms, which refers to the phenomenon that the algorithm returns solutions with good generalization despite the presence of solutions with poor generalization due to the overparametrized models. We first give a rigorous justification of why convolutional networks are more sample-efficient than fully-connected networks. Then we provide theoretical justification for the empirical observation that deep linear networks, including matrix factorization, trained by gradient descent from small initialization implicitly bias to low-rank solutions. We also identify a condition when gradient descent with reparametrization is equivalent to mirror descent which can be used to understand implicit bias of non-linear models and recovers several previous results. We further show gradient descent has an implicit bias for `flatter' solutions when having certain gradient noise or its learning rate is larger than 2 over sharpness of loss. Louis Riehl Graduate Administrator Computer Science Department, CS213 Princeton University (609) 258-8014

1 0