- talks - lists.cs.princeton.edu

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

Ankit Goyal will present his FPO "Towards Geometric Intelligence" on Monday, August 15, 2022 at 10:00 AM in Friend Center 202 and Zoom.
by Louis W. Riehl 08 Aug '22

08 Aug '22

Ankit Goyal will present his FPO "Towards Geometric Intelligence" on Monday, August 15, 2022 at 10:00 AM in Friend Center 202 and Zoom. Location: Zoom link: https://princeton.zoom.us/my/agoyal The members of Ankit’s committee are as follows: Examiners: Jia Deng (Adviser), Szymon Rusinkiewicz, Danqi Chen Readers: Karthik Narasimhan, Vladlen Koltun (Apple) 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: Geometric intelligence is the aspect of human intelligence that relates to perceiving, communicating and reasoning about geometries. Artificial agents like robots must possess geometric intelligence to operate in unconstrained human environments and collaborate with humans in day-to-day life. Geometric intelligence is an umbrella term that encompasses many abilities. In this thesis, we focus on three crucial abilities – first, the ability to see or perceive geometry; second, the ability to communicate about geometry and space and third, the ability to reason and plan about geometries. For studying the ability to perceive geometry, we pursue two efforts. In one effort, we build a system, called IFOR that recognizes the geometric difference between two scenes and rearrange one scene into another. IFOR can handle unseen objects and transfer to the real world while being trained only on synthetic data. In another effort, we revisit the literature on perceiving objects from point clouds and uncover two surprising results. First, we show that auxiliary factors that are independent of network architecture explain most of the performance improvement. Second, we show that a simple view-based baseline outperforms sophisticated state-of-the-art methods. For studying the ability to communicate about geometry, we focus on recognizing spatial relations, which are the atomic elements used to communicate about geometric arrangements. We find that existing datasets are insufficient as they lack largescale, high-quality 3D ground truth information, which is critical for learning spatial relations. We fill this gap by constructing Rel3D. Finally, for studying the ability to reason about geometry, we pursue two efforts. In one effort, we study the problem of geometric reasoning in the context of question-answering. We introduce Dynamic Spatial Memory Network (DSMN), a new deep network architecture designed for answering questions that admit latent visual representations. In another effort, we explore the problem of geometric planning which requires simultaneous reasoning about geometries and planning. We find that existing benchmarks are insufficient for the problem. We propose PackIt – a virtual environment for geometric planning. We benchmark various baselines on PackIt and find that learning could be a viable way to gain geometric planning skills. Louis Riehl Graduate Administrator Computer Science Department, CS213 Princeton University (609) 258-8014

1 0

Princeton AI Club speaker: Prof. Max Welling, August 15th at 10 AM ET
by Emily C. Lawrence 04 Aug '22

04 Aug '22

The Princeton AI Club is thrilled to announce its next guest speaker – [ https://staff.fnwi.uva.nl/m.welling/ | Max Welling ] from the University of Amsterdam! His talk will focus on how we can apply machine learning to scientific discovery If you want to about new opportunities to apply machine learning to scientific discovery, sign-up for his virtual talk [ https://forms.gle/ms5hewGsMc8SUgaL7 | here ] , and join us on August 15 th , at 10 am Eastern Time (ET) ! The full details are below. Speaker: Prof. Max Welling (University of Amsterdam) Title: Deep Learning and the Natural Sciences: A Perfect Marriage? Date: Monday, August 15 at 4:00PM CEST (which is 10:00 AM EST.) Abstract: Artificial Intelligence has always had a deep connection with the natural sciences. Artificial neural networks were first conceived as abstract ions of biological neural networks, as do many subsequent algorithms such as reinforcement learning. Neural networks have also been inspired by many concepts in physics, for example Hopfield networks are modelled after Ising models. More recently, concepts from physics such as symmetries and the renormalization group have also been incorporated in the design of novel deep architectures. With these new tools, deep learning has completely revolutionized fields such as Natural Language Processing, and Computer Vision. Also in the reverse direction, machine learning has contributed tremendously to the analysis of data from experiments, such as particle collision events at CERN. However, recently, a new scientific paradigm is emerging where deep learning models are accelerating and improving scientific simulations, dubbed by some the fifth scientific paradigm. In this talk I will first give an introduction to the opportunities that lie ahead of us to apply machine learning to scientific discovery. In the second half of the talk, I will talk in more technical detail about symmetries in deep learning and our latest ideas on how to model the hidden layers of a deep neural network as a field described by a partial differential equation. Amazingly, this field-view of the hidden layers of an artificial cortex can be extended to a quantum field and mapped to an optical quantum computer design. How useful this intellectual exercise is in terms of practical quantum computers I will leave to the imagination of the audience. Bio: Prof. Dr. Max Welling is a research chair in Machine Learning at the University of Amsterdam and a Distinguished Scientist at MSR. He is a fellow at the Canadian Institute for Advanced Research (CIFAR) and the European Lab for Learning and Intelligent Systems (ELLIS) where he also serves on the founding board. His previous appointm z ents include VP at Qualcomm Technologies, professor at UC Irvine, postdoc at U. Toronto and UCL under supervision of prof. Geoffrey Hinton, and postdoc at Caltech under supervision of prof. Pietro Perona. He finished his PhD in theoretical high energy physics under supervision of Nobel laureate prof. Gerard ‘t Hooft. Max Welling has served as associate editor in chief of IEEE TPAMI from 2011-2015, he serves on the advisory board of the Neurips foundation since 2015 and has been program chair and general chair of Neurips in 2013 and 2014 respectively. He was also program chair of AISTATS in 2009 and ECCV in 2016 and general chair of MIDL 2018. Max Welling is recipient of the ECCV Koenderink Prize in 2010 and the ICML Test of Time award in 2021. He directs the Amsterdam Machine Learning Lab (AMLAB) and co-directs the Qualcomm-UvA deep learning lab (QUVA) and the Bosch-UvA Deep Learning lab (DELTA).

1 0

Zhuqi Li will present his Pre FPO "Cross-layer optimization for video delivery on Wireless Networks" on Wednesday, August 10, 2022 at 3pm (EST) via Zoom.
by Nicki Mahler 03 Aug '22

03 Aug '22

Zhuqi Li will present his Pre FPO "Cross-layer optimization for video delivery on Wireless Networks" on Wednesday, August 10, 2022 at 3pm (EST) via Zoom. Zoom link: [ https://princeton.zoom.us/j/6871275896 | https://princeton.zoom.us/j/6871275896 ] The members of his committee are as follows: Advisor: Kyle Jamieson Readers: Victor Bahl (Microsoft), Amit Levy Examiners: Ravi Netravali, Jennifer Rexford All are welcome to attend. Abstract: A dominating portion of today's wireless network traffic is occupied by video content delivery for both video streaming and live video analytic purposes. The dynamic nature of wireless network presents significant challenges to both applications. On the one hand, the available throughput of the wireless network varies across time due to the change of the wireless channel and competing flows. On the other hand, the endpoints need to decide the video bitrate and routing selection without knowing the future throughput. In this talk, we will introduce three systems that optimize wireless video delivery from physical layer, network layer, and application layer to solve the challenges. We present the design and implementation of the following systems: (1) LAIA: a programmable wireless environment that allows rapid adjustments of the wireless channel to smooth the throughput variation. LAIA uses a large array of inexpensive antenna (LAIA) elements and includes algorithms to configure LAIA elements in real-time. (2) Spider: a multi-hop, millimeter-wave (mmWave) wireless relay network design to maximize end-to-end video analytic accuracy for a large camera cluster. Spider integrates a low-latency Wi-Fi control plane with a mmWave relay data plane, allowing agile re-routing around blockages. Spider proposes a novel video bit-rate allocation algorithm coupled with a scalable routing algorithm that works hand-in-hand toward the application-level objective of maximizing video analytics accuracy, rather than simply maximizing data throughput. (3) Dashlet: a system tailored for the high quality of experience in short video streaming applications. Dashlet proposes a novel out-of-order video chunk pre-buffering mechanism that leverages users’ swipe statistics as well as network throughput history to determine the pre-buffering order and bitrate.

1 0

Seminar Series on Security & Privacy in Machine Learning
by Emily C. Lawrence 02 Aug '22

02 Aug '22

Challenges and Opportunities in Security & Privacy in Machine Learning Weekly talks starting Tuesday, June 7 About the seminar series The motivation for the seminar is to build a platform to discuss and disseminate the progress made by the community in solving some of the core challenges. We intend to host weekly talks from leading researchers in both academia and industry. Each session will be split into a talk (40 mins) followed by a Q&A + short discussion session (20 mins). Timing : Every Tuesday at 1pm Eastern Time (Virtual talks) Website: [ https://vsehwag.github.io/SPML_seminar/ | https://vsehwag.github.io/SPML_seminar/ ] Mailing list: [ https://groups.google.com/forum/#!forum/spml-seminars/join | Link to mailing list ] Calendar: [ https://calendar.google.com/calendar/u/0?cid=N2FwbTVxYzJsOGM2bXBiNGY4am1oMj… | Link to calendar ] You can find all additional details on the website. If you are interested, we recommend signing up for the mailing list and sync the calendar to stay up to date with the seminar schedule. -- Vikash Sehwag PhD Candidate Princeton University [ https://vsehwag.github.io/ | https://vsehwag.github.io ]

1 8

Happening today! Our next guest speaker is Dr. Naila Murray from Facebook AI Research!
by Emily C. Lawrence 01 Aug '22

01 Aug '22

Happening today! The Princeton AI Club is excited to announce its next guest speaker – [ https://twitter.com/nailamurray | Dr. Naila Murray ] from Facebook AI Research! Her talk will focus on how to train deep vision models in low-data regimes If you want to hear all about how to train your network with little data, sign-up for her virtual talk [ https://forms.gle/fThq5gxa6Yzx1dkh8 | here ] , and join us on August 1 st , at 11 am Eastern Time (ET) ! The full details are below. Speaker: Dr. Naila Murray (Facebook AI Research) Title: Training deep vision models in low-data regimes Date: Monday, August 1st 2022, 11:00 AM ET Abstract: Large-scale vision models, trained on vast quantities of often unlabelled data, give state-of-the-art performance when fine-tuned for a wide variety of downstream tasks. For these downstream tasks, the amount of training data available is usually orders of magnitude less than what was used during pre-training. And in many cases, there is little to no labelled data available. In this talk, I discuss approaches to fine-tuning pre-trained models for vision tasks in low-data regimes. In all cases, which cover retrieval, video segmentation, and object detection, I show that incorporating domain-specific and modality-specific inductive biases lead to improved model performance when training data is limited. Bio: Naila Murray obtained a BSE in electrical engineering from Princeton University in 2007. In 2012, she received her Ph.D. from the Universitat Autonoma de Barcelona, in affiliation with the Computer Vision Center. She joined Xerox Research Centre Europe in 2013 as a research scientist in the computer vision team, working on topics including fine-grained visual categorization, image retrieval and visual attention. From 2015 to 2019 she led the computer vision team at Xerox Research Centre Europe, and continued to serve in this role after its acquisition and transition to becoming NAVER LABS Europe. In 2019, she became the director of science at NAVER LABS Europe. In 2020, she joined Facebook AI Research where she is a senior research engineering manager for EMEA. She has served as area chair for ICLR 2018, ICCV 2019, ICLR 2019, CVPR 2020, ECCV 2020, and program chair for ICLR 2021. Her current research interests include few-shot learning and domain adaptation.

1 0

Princeton AI Club: Dr. Naila Murray from Facebook AI Research, August 1st 2022, at 11 AM ET
by Emily C. Lawrence 22 Jul '22

22 Jul '22

Reminder for our next guest speaker! The Princeton AI Club is excited to announce its next guest speaker – [ https://twitter.com/nailamurray | Dr. Naila Murray ] from Facebook AI Research! Her talk will focus on how to train deep vision models in low-data regimes If you want to hear all about how to train your network with little data, sign-up for her virtual talk [ https://forms.gle/fThq5gxa6Yzx1dkh8 | here ] , and join us on August 1 st , at 11 am Eastern Time (ET) ! The full details are below. Speaker: Dr. Naila Murray (Facebook AI Research) Title: Training deep vision models in low-data regimes Date: Monday, August 1st 2022, 11:00 AM ET Abstract: Large-scale vision models, trained on vast quantities of often unlabelled data, give state-of-the-art performance when fine-tuned for a wide variety of downstream tasks. For these downstream tasks, the amount of training data available is usually orders of magnitude less than what was used during pre-training. And in many cases, there is little to no labelled data available. In this talk, I discuss approaches to fine-tuning pre-trained models for vision tasks in low-data regimes. In all cases, which cover retrieval, video segmentation, and object detection, I show that incorporating domain-specific and modality-specific inductive biases lead to improved model performance when training data is limited. Bio: Naila Murray obtained a BSE in electrical engineering from Princeton University in 2007. In 2012, she received her Ph.D. from the Universitat Autonoma de Barcelona, in affiliation with the Computer Vision Center. She joined Xerox Research Centre Europe in 2013 as a research scientist in the computer vision team, working on topics including fine-grained visual categorization, image retrieval and visual attention. From 2015 to 2019 she led the computer vision team at Xerox Research Centre Europe, and continued to serve in this role after its acquisition and transition to becoming NAVER LABS Europe. In 2019, she became the director of science at NAVER LABS Europe. In 2020, she joined Facebook AI Research where she is a senior research engineering manager for EMEA. She has served as area chair for ICLR 2018, ICCV 2019, ICLR 2019, CVPR 2020, ECCV 2020, and program chair for ICLR 2021. Her current research interests include few-shot learning and domain adaptation.

1 0

Kaiyu Yang will present his FPO "Neurosymbolic Machine Learning for Reasoning" on Thursday, July 28, 2022 at 3:00 PM in Friend 125 and Zoom.
by Louis W. Riehl 21 Jul '22

21 Jul '22

Kaiyu Yang will present his FPO "Neurosymbolic Machine Learning for Reasoning" on Thursday, July 28, 2022 at 3:00 PM in Friend 125 and Zoom. Location: Zoom link: [ https://princeton.zoom.us/j/2731344683 | https://princeton.zoom.us/j/2731344683 ] The members of Kaiyu’s committee are as follows: Examiners: Jia Deng (Adviser), Olga Russakovsky, Danqi Chen Readers: Karthik Narasimhan, Mayur Naik (University of Pennsylvania) 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: Machine learning has made significant progress in the past decade. Its most successful paradigm, deep neural networks, consists of layers of continuous representation whose parameters are optimized on massive datasets via gradient descent. Deep neural networks have achieved superior performance on numerous tasks, such as object recognition, language understanding, and autonomous driving. However, they still struggle with reasoning tasks, which often require manipulating symbols and chaining multiple steps compositionally, e.g., solving math equations or writing computer programs. In this dissertation, we aim to bridge this gap and teach machines to reason—in ways that are precise, systematic, interpretable, and robust to ambiguity in real-world environments. To that end, we take a neurosymbolic approach combining the complementary strengths of machine learning and symbolic reasoning. Symbolic reasoning is precise and generalizes systematically. But it has been restricted to domains amenable to rigid formalization. In contrast, predominant machine learning methods are flexible but notoriously uninterpretable, data-hungry, and incapable of generalizing outside the training distribution. Integrating the strengths of both approaches is essential for building flexible reasoning machines with precise and systematic generalization. Concretely, this dissertation addresses neurosymbolic reasoning from two angles. First, we apply machine learning to tasks related to symbolic reasoning, such as automated theorem proving (Chapter 2). Second, we introduce inductive biases inspired by symbolic reasoning into machine learning models to improve their interpretability, generalization, and data efficiency (Chapter 3 and Chapter 4). Our results highlight the benefits of (1) neurosymbolic model architectures, (2) reasoning at a suitable level of abstraction, and (3) an explicit, compositional representation of reasoning, such as symbolic proofs. Louis Riehl Graduate Administrator Computer Science Department, CS213 Princeton University (609) 258-8014

1 0

Xingyuan Sun will present his Pre-FPO "Using Machine Learning to Improve Fiber-Reinforced 3D Printed Objects" on Thursday, July 28, 2022 at 11:00 AM in CS 402 and Zoom.
by Louis W. Riehl 21 Jul '22

21 Jul '22

Xingyuan Sun will present his Pre-FPO "Using Machine Learning to Improve Fiber-Reinforced 3D Printed Objects" on Thursday, July 28, 2022 at 11:00 AM in CS 402 and Zoom. Zoom link: [ https://princeton.zoom.us/j/95907350060 | https://princeton.zoom.us/j/95907350060 ] Committee Members: Ryan P. Adams (Examiner, adviser), Szymon Rusinkiewicz (Examiner, adviser), Olga Russakovsky (Examiner), Sigrid Adriaenssens (reader), Felix Heide (reader). All are welcome to attend. Title: Using Machine Learning to Improve Fiber-Reinforced 3D Printed Objects Abstract: This dissertation studies fiber layout planning given a shape and external loads using machine learning, improving the object’s stiffness. However, simply planning fiber paths is not sufficient. Since fibers are stiff, their shape is smoothed after extrusion. We thus need another planning algorithm to generate an extruder path given any desired fiber path. To solve this task, we learn a two-stage neural network that we may consider an autoencoder. A decoder is first learned as a differentiable surrogate of the printing process, taking an extruder path and producing a resulting fiber path. An encoder is then trained to produce an extruder path given a desired fiber path, using the trained decoder to evaluate the quality of the planned extruder path. We compare this approach with direct supervised learning of the extruder path planning problem and direct optimization of the extruder path. Besides, we also study our algorithm in another case study on a soft robot, showing our approach can be applied to design problems with a similar structure. Results from simulation and real experiments show our algorithm generates higher quality solutions than direct supervised learning and solutions with competitive quality with direct optimization but greatly reduced running time. To plan fiber paths, we formalize the task as an optimization problem, solving the fiber layout to maximize the object’s stiffness with regularizers forcing paths to be feasible. We initialize every fiber path by iteratively using finite element analysis to calculate the stress field of the object and a greedy algorithm to “walk” on the field. We then use the adjoint method to calculate the gradient of the regularized stiffness with respect to the fiber layout, and optimize the layout using a gradient-based optimizer. We compare our algorithm with three baselines, including concentric fiber rings generated by Eiger, a leading digital manufacturing software, and two others producing greedy fiber paths on the stress field and a fiber orientation field calculated by smoothing the stress field, respectively. Experiments on simulation and real printings show objects with fiber paths from our algorithm improve the Pareto front, achieving larger stiffness while using less fiber. Louis Riehl Graduate Administrator Computer Science Department, CS213 Princeton University (609) 258-8014

1 0