- talks - lists.cs.princeton.edu

Andrew Or will present his Pre FPO "Abstracting systems challenges from distributed deep learning" on Wednesday, March 17, 2021 at 1pm via Zoom.
by Nicki Mahler 10 Mar '21

10 Mar '21

Andrew Or will present his Pre FPO " Abstracting systems challenges from distributed deep learning" on Wednesday, March 17, 2021 at 1pm via Zoom. Zoom link: [ https://princeton.zoom.us/j/91618635317?pwd=cyt5WmdFQ2xaWmNGb2U0bTNlZXBYZz09 | https://princeton.zoom.us/j/91618635317?pwd=cyt5WmdFQ2xaWmNGb2U0bTNlZXBYZz09 ] Committee: Michael J. Freedman (advisor, examiner) Kai Li (examiner) Wyatt Lloyd (examiner) Amit Levy (reader) Ravi Netravali (reader) Talk abstract follows below. All are welcome to attend. Abstract: State-of-the-art distributed deep learning systems tightly couple model execution with the underlying hardware. Existing frameworks assume resource allocations must be fixed throughout the lifetime of a job, often leading to inefficiency in the usage of cluster resources. Further, model hyperparameters need to be retuned across different hardware configurations in order to achieve the same training result. In a world where the scale of deep learning workloads is growing rapidly, this requirement poses significant barriers to experimentation on smaller test beds and reproducing results across different hardware. In this thesis, we demonstrate that the above assumptions are not fundamental to distributed deep learning, and we resolve these limitations by proposing two systems built on top of TensorFlow. The first is an autoscaling engine that, through trial-and-error, automatically determines the most resource-efficient hardware configuration for a given job. We propose pluggable heuristics tailored for deep learning workloads that incrementally guide the system towards such a configuration. Instead of repeatedly stopping the job and restarting it from checkpoints, which can lead to GPUs or TPUs going idle for minutes every time, our system adjusts the all-reduce membership dynamically in between batches without interrupting the job. The second system is VirtualFlow, which leverages a novel abstraction between the model and the underlying hardware called virtual node processing. From the perspective of the model, virtual nodes, instead of physical devices (GPUs or TPUs), perform the computation. When multiple virtual nodes are mapped to a physical device, they are processed sequentially on that device. This representation allows users to train a model using the same batch size (and other hyperparameters) across different hardware, thereby preserving application-level semantics while hiding hardware-level details from the user. Using this technique, VirtualFlow also enables a variety of new important use cases that improve resource efficiency, including resource elasticity and heterogeneous training (using different types of GPUs in the same job).

1 0

CS Colloquium Speaker: Deepak Narayanan, today at 4:30pm EST
by Emily C. Lawrence 10 Mar '21

10 Mar '21

CS Colloquium Speaker Deepak Narayanan, Stanford University Host: Kyle Jamieson Date: Wednesday, March 10 Time: 4:30pm EST Event page: [ https://www.cs.princeton.edu/events/26048 | https://www.cs.princeton.edu/events/26048 ] Register for Zoom webinar here: [ https://princeton.zoom.us/webinar/register/WN_Gry8g80-S06bARgqKzko_Q | https://princeton.zoom.us/webinar/register/WN_Gry8g80-S06bARgqKzko_Q ] Title: Resource-Efficient Execution for Deep Learning Abstract: Deep Learning models have enabled state-of-the-art results across a broad range of applications; however, training these models is extremely time- and resource-intensive, taking weeks on clusters with thousands of expensive accelerators in the extreme case. In this talk, I will describe two systems that improve the resource efficiency of model training. The first system, PipeDream, proposes the use of pipelining to accelerate distributed training. Pipeline parallelism facilitates model training with lower communication overhead than previous methods while still ensuring high compute resource utilization. Pipeline parallelism also enables the efficient training of large models that do not fit on a single worker. Pipeline parallelism is being used at Facebook, Microsoft, OpenAI, and Nvidia for efficient large-scale model training. The second system, Gavel, determines how resources in a shared cluster with heterogeneous compute resources (e.g., different types of hardware accelerators) should be partitioned among different users to optimize objectives specified over multiple training jobs. Gavel can improve various scheduling objectives, such as average completion time, makespan, or cloud computing resource cost, by up to 3.5x. I will conclude the talk with discussion on future directions for optimizing Machine Learning systems. Bio: Deepak Narayanan is a final-year PhD student at Stanford University advised by Prof. Matei Zaharia. He is interested in designing and building software to improve the runtime performance and efficiency of emerging machine learning and data analytics workloads on modern hardware. His work is supported by a NSF graduate fellowship.

1 0

Reminder: Deepak Narayanan, tomorrow at 4:30
by Emily C. Lawrence 09 Mar '21

09 Mar '21

CS Colloquium Speaker Deepak Narayanan, Stanford University Host: Kyle Jamieson Date: Wednesday, March 10 Time: 4:30pm EST Event page: [ https://www.cs.princeton.edu/events/26048 | https://www.cs.princeton.edu/events/26048 ] Register for Zoom webinar here: [ https://princeton.zoom.us/webinar/register/WN_Gry8g80-S06bARgqKzko_Q | https://princeton.zoom.us/webinar/register/WN_Gry8g80-S06bARgqKzko_Q ] Title: Resource-Efficient Execution for Deep Learning Abstract: Deep Learning models have enabled state-of-the-art results across a broad range of applications; however, training these models is extremely time- and resource-intensive, taking weeks on clusters with thousands of expensive accelerators in the extreme case. In this talk, I will describe two systems that improve the resource efficiency of model training. The first system, PipeDream, proposes the use of pipelining to accelerate distributed training. Pipeline parallelism facilitates model training with lower communication overhead than previous methods while still ensuring high compute resource utilization. Pipeline parallelism also enables the efficient training of large models that do not fit on a single worker. Pipeline parallelism is being used at Facebook, Microsoft, OpenAI, and Nvidia for efficient large-scale model training. The second system, Gavel, determines how resources in a shared cluster with heterogeneous compute resources (e.g., different types of hardware accelerators) should be partitioned among different users to optimize objectives specified over multiple training jobs. Gavel can improve various scheduling objectives, such as average completion time, makespan, or cloud computing resource cost, by up to 3.5x. I will conclude the talk with discussion on future directions for optimizing Machine Learning systems. Bio: Deepak Narayanan is a final-year PhD student at Stanford University advised by Prof. Matei Zaharia. He is interested in designing and building software to improve the runtime performance and efficiency of emerging machine learning and data analytics workloads on modern hardware. His work is supported by a NSF graduate fellowship.

1 0

TOMORROW 3/9/2021 ECE Seminar - Maria Apostolaki, ETH
by Emily C. Lawrence 08 Mar '21

08 Mar '21

ECE SEMINAR Speaker: Maria Apostolaki, ETH Title: Building secure distributed systems atop the insecure Internet Day: Tuesday, March 9, 2021 Time: 4:30 pm Room: Please register at [ https://princeton.zoom.us/meeting/register/tJUlc-2oqTwiHdZVqHbVqzMkeIX1y6v7… | https://princeton.zoom.us/meeting/register/tJUlc-2oqTwiHdZVqHbVqzMkeIX1y6v7… ] Host: Prateek Mittal Abstract: Distributed systems are increasingly important for our everyday life, allowing for high performance, fault tolerance, and flexibility. Many of these systems nowadays rely on the inherently insecure Internet infrastructure. Surely though, they should have been designed to take this into account...? In this talk, I will answer negatively to this question using a concrete example: public blockchain systems such as Bitcoin. These are novel distributed systems that are designed according to stringent failure models. In this context, I will explain how an adversary controlling pieces of Internet infrastructure can practically compromise: (i) Bitcoin's consensus protocol (by partitioning the network); (ii) Bitcoin's anonymity guarantees (by mapping pseudonyms to real-world identities); and (iii) Bitcoin's availability (by eclipsing clients). While these attacks are worrying, I will also introduce practical and effective defenses to counter them both at the network and the application layer. Beyond Bitcoin, this work teaches essential lessons for distributed-system design. Bio: Maria Apostolaki is a PhD Student at ETH Zurich advised by Laurent Vanbever. During her studies, she has been a visiting student at MIT (2019) and a research intern at Microsoft Research (2018) and Google (2017). Before joining ETH, she earned her diploma in Electrical and Computer Engineering at the National Technical University of Athens, Greece. Her research focuses on building secure, performant, and deployable networked systems using both hardware and software. Her research led to discovering significant vulnerabilities in the Bitcoin system and spearheaded changes to the Bitcoin codebase. Maria’s work received widespread media coverage and was awarded an Applied Networking Research Prize by IETF.

1 0

CS Colloquium Speaker: Amy Zhang, today at 4:30pm
by Emily C. Lawrence 08 Mar '21

08 Mar '21

CS Colloquium Speaker Amy Zhang, McGill University and the Mila Institute Date: Monday, March 8 Time: 4:30pm EST Host: Tom Griffiths Event page: [ https://www.cs.princeton.edu/events/26041 | https://www.cs.princeton.edu/events/26041 ] Register for Zoom webinar here: [ https://princeton.zoom.us/webinar/register/WN_yb-TRAQJSwyBTo7RSDfhUA | https://princeton.zoom.us/webinar/register/WN_yb-TRAQJSwyBTo7RSDfhUA ] Title: Exploiting Latent Structure and Bisimulation Metrics for Better Generalization in Reinforcement Learning Abstract: The advent of deep learning has shepherded unprecedented progress in various fields of machine learning. Despite recent advances in deep reinforcement learning (RL) algorithms, however, there is no method today that exhibits anywhere near the generalization that we have seen in computer vision and NLP. Indeed, one might ask whether deep RL algorithms are even capable of the kind of generalization that is needed for open-world environments. This challenge is fundamental and will not be solved with incremental algorithmic advances. In this talk, we propose to incorporate different assumptions that better reflect the real world and allow the design of novel algorithms with theoretical guarantees to address this fundamental problem. We first present how state abstractions can accelerate reinforcement learning from rich observations, such as images, without relying either on domain knowledge or pixel-reconstruction. Our goal is to learn state abstractions that both provide for effective downstream control and invariance to task-irrelevant details. We use bisimulation metrics to quantify behavioral similarity between states, and learn robust latent representations which encode only the task-relevant information from observations. We provide theoretical guarantees for the learned approximate abstraction and extend this notion to families of tasks with varying dynamics. Bio: I am a final year PhD candidate at McGill University and the Mila Institute, co-supervised by Profs. Joelle Pineau and Doina Precup. I am also a researcher at Facebook AI Research. My work focuses on bridging theory and practice through learning approximate state abstractions and learning representations for generalization in reinforcement learning. I previously obtained an M.Eng. in EECS and dual B.Sci. degrees in Mathematics and EECS from MIT.

1 0

Reminder: Amy Zhang, tomorrow at 4:30pm
by Emily C. Lawrence 07 Mar '21

07 Mar '21

CS Colloquium Speaker Amy Zhang, McGill University and the Mila Institute Date: Monday, March 8 Time: 4:30pm EST Host: Tom Griffiths Event page: [ https://www.cs.princeton.edu/events/26041 | https://www.cs.princeton.edu/events/26041 ] Register for Zoom webinar here: [ https://princeton.zoom.us/webinar/register/WN_yb-TRAQJSwyBTo7RSDfhUA | https://princeton.zoom.us/webinar/register/WN_yb-TRAQJSwyBTo7RSDfhUA ] Title: Exploiting Latent Structure and Bisimulation Metrics for Better Generalization in Reinforcement Learning Abstract: The advent of deep learning has shepherded unprecedented progress in various fields of machine learning. Despite recent advances in deep reinforcement learning (RL) algorithms, however, there is no method today that exhibits anywhere near the generalization that we have seen in computer vision and NLP. Indeed, one might ask whether deep RL algorithms are even capable of the kind of generalization that is needed for open-world environments. This challenge is fundamental and will not be solved with incremental algorithmic advances. In this talk, we propose to incorporate different assumptions that better reflect the real world and allow the design of novel algorithms with theoretical guarantees to address this fundamental problem. We first present how state abstractions can accelerate reinforcement learning from rich observations, such as images, without relying either on domain knowledge or pixel-reconstruction. Our goal is to learn state abstractions that both provide for effective downstream control and invariance to task-irrelevant details. We use bisimulation metrics to quantify behavioral similarity between states, and learn robust latent representations which encode only the task-relevant information from observations. We provide theoretical guarantees for the learned approximate abstraction and extend this notion to families of tasks with varying dynamics. Bio: I am a final year PhD candidate at McGill University and the Mila Institute, co-supervised by Profs. Joelle Pineau and Doina Precup. I am also a researcher at Facebook AI Research. My work focuses on bridging theory and practice through learning approximate state abstractions and learning representations for generalization in reinforcement learning. I previously obtained an M.Eng. in EECS and dual B.Sci. degrees in Mathematics and EECS from MIT.

1 0

Riley Simmons-Edler will present his Pre FPO "Overcoming Sampling and Exploration Challenges in Deep Reinforcement Learning" on Friday, March 12, 2021 at 4pm via Zoom
by Nicki Mahler 05 Mar '21

05 Mar '21

Riley Simmons-Edler will present his Pre FPO "Overcoming Sampling and Exploration Challenges in Deep Reinforcement Learning" on Friday, March 12, 2021 at 4pm via Zoom [ https://princeton.zoom.us/j/92165806242?pwd=VWhvWW1DZXpqWUsvY0pVclE1eklVZz09 | https://princeton.zoom.us/j/92165806242?pwd=VWhvWW1DZXpqWUsvY0pVclE1eklVZz09 ] Committee: Sebastian Seung (advisor, examiner) Olga Russokovsky (examiner) Szymon Rusinkiewicz (examiner) Daniel Lee (external- Cornell, reader) Karthik Narasimhan (reader) Talk abstract follows below. All are welcome to attend. Abstract: The combination of deep neural networks with the algorithms and formalisms of reinforcement learning (RL) shows great promise for solving otherwise intractable learning tasks. However, practical applications of deep reinforcement learning remain scarce. The outstanding challenges of deep RL can be broadly grouped into two categories, broadly described as "What to learn from experiences?" and "What experiences to learn from?" In this thesis, I will describe the work I have done to address the second category, which is historically the less-studied domain. Specifically, I address problems of sampling- whether actions, states, or trajectories- which contain information sufficient for learning a task. I examine this challenge at three levels of algorithm design and task complexity, from algorithmic submodules to combinations of algorithms to hybrid algorithms that violate common RL conventions. First, I will present my work on stable and efficient sampling of actions that optimize a Q-function of continuous-valued actions. By combining a sample-based optimizer with neural network approximation, it is possible to obtain both stability in training as well as computational efficiency and precision in inference. Second, I will present my work on reward-aware exploration, the discovery of highly-rewarding states in tasks where commonly-used sampling methods are insufficient. A teacher "exploration" agent can discover states and trajectories which maximize the amount a student "exploitation" agent learns on those experiences, and can enable the student agent to solve hard tasks which are otherwise impossible for it. Third, I will present my work combining reinforcement learning with heuristic search, for use in task domains where the transition model is known, but where the combinatorics of the state space are intractable for traditional search algorithms. I show that by combining deep Q-learning with a best-first tree search algorithm, it is possible to find solutions to simple program synthesis problems with dramatically fewer samples than common search algorithms require. Lastly, I will conclude with a summary of the major takeaways of this work, and discuss potential extensions and future directions for the efficient sampling of useful experiences in RL.

1 0

CS Colloquium Speaker: Huacheng Yu today at 12:30pm EST
by Emily C. Lawrence 04 Mar '21

04 Mar '21

CS Colloquium Speaker Huacheng Yu, Princeton University Host: Zeev Dvir Date: Thursday, March 4 Time: 12:30pm EST Event page: [ https://www.cs.princeton.edu/events/26051 | https://www.cs.princeton.edu/events/26051 ] Register for Zoom webinar here: [ https://princeton.zoom.us/webinar/register/WN_pNvnCKFHS_qoXLEEOSMOwA | https://princeton.zoom.us/webinar/register/WN_pNvnCKFHS_qoXLEEOSMOwA ] Title: Data Structures and Algorithms in Sublinear Computation Abstract: Sublinear algorithms are sustainable under the exponential increase of data volume and processing speed. Such algorithms use a sublinear amount of resources, e.g., spending time, space, or communication that is asymptotically smaller than the input data size. Typical examples include data structures, which compute query functions of the data in sublinear time, and streaming algorithms, which make one pass over massive data streams while maintaining a sublinear-sized memory. In this talk, I will give an overview of my work in sublinear computation, focusing on succinct data structures and distributed graph sketching algorithms. I will first discuss my work on a nearly optimal data structure for the dictionary problem, for which the textbook solution uses hash tables. Then, I will talk about detecting the connectivity of graphs using distributed sketching, and my recent work showing the optimality of a well-known sketching algorithm (the AGM sketch). I will conclude the talk with discussion on future directions and my other work in theoretical computer science. Bio: Huacheng Yu is an associate research scholar in the Department of Computer Science at Princeton University. His research interests include data structures and streaming algorithms, and other directions in theoretical computer science such as communication complexity and graph algorithms. Prior to Princeton, Huacheng was a postdoctoral researcher at Harvard University hosted by Jelani Nelson and Madhu Sudan. He received his Ph.D. from Stanford University (advised by Ryan Williams and Omer Reingold) and B.Eng from Tsinghua University, both in Computer Science.

1 0

CS Colloquium Speaker: Deepak Narayanan, Wednesday, March 10 at 4:30
by Emily C. Lawrence 03 Mar '21

03 Mar '21

CS Colloquium Speaker Deepak Narayanan, Stanford University Host: Kyle Jamieson Date: Wednesday, March 10 Time: 4:30pm EST Event page: [ https://www.cs.princeton.edu/events/26048 | https://www.cs.princeton.edu/events/26048 ] Register for Zoom webinar here: [ https://princeton.zoom.us/webinar/register/WN_Gry8g80-S06bARgqKzko_Q | https://princeton.zoom.us/webinar/register/WN_Gry8g80-S06bARgqKzko_Q ] Title: Resource-Efficient Execution for Deep Learning Abstract: Deep Learning models have enabled state-of-the-art results across a broad range of applications; however, training these models is extremely time- and resource-intensive, taking weeks on clusters with thousands of expensive accelerators in the extreme case. In this talk, I will describe two systems that improve the resource efficiency of model training. The first system, PipeDream, proposes the use of pipelining to accelerate distributed training. Pipeline parallelism facilitates model training with lower communication overhead than previous methods while still ensuring high compute resource utilization. Pipeline parallelism also enables the efficient training of large models that do not fit on a single worker. Pipeline parallelism is being used at Facebook, Microsoft, OpenAI, and Nvidia for efficient large-scale model training. The second system, Gavel, determines how resources in a shared cluster with heterogeneous compute resources (e.g., different types of hardware accelerators) should be partitioned among different users to optimize objectives specified over multiple training jobs. Gavel can improve various scheduling objectives, such as average completion time, makespan, or cloud computing resource cost, by up to 3.5x. I will conclude the talk with discussion on future directions for optimizing Machine Learning systems. Bio: Deepak Narayanan is a final-year PhD student at Stanford University advised by Prof. Matei Zaharia. He is interested in designing and building software to improve the runtime performance and efficiency of emerging machine learning and data analytics workloads on modern hardware. His work is supported by a NSF graduate fellowship.

1 0

Reminder: Huacheng Yu tomorrow at 12:30pm EST
by Emily C. Lawrence 03 Mar '21

03 Mar '21

CS Colloquium Speaker Huacheng Yu, Princeton University Host: Zeev Dvir Date: Thursday, March 4 Time: 12:30pm EST Event page: [ https://www.cs.princeton.edu/events/26051 | https://www.cs.princeton.edu/events/26051 ] Register for Zoom webinar here: [ https://princeton.zoom.us/webinar/register/WN_pNvnCKFHS_qoXLEEOSMOwA | https://princeton.zoom.us/webinar/register/WN_pNvnCKFHS_qoXLEEOSMOwA ] Title: Data Structures and Algorithms in Sublinear Computation Abstract: Sublinear algorithms are sustainable under the exponential increase of data volume and processing speed. Such algorithms use a sublinear amount of resources, e.g., spending time, space, or communication that is asymptotically smaller than the input data size. Typical examples include data structures, which compute query functions of the data in sublinear time, and streaming algorithms, which make one pass over massive data streams while maintaining a sublinear-sized memory. In this talk, I will give an overview of my work in sublinear computation, focusing on succinct data structures and distributed graph sketching algorithms. I will first discuss my work on a nearly optimal data structure for the dictionary problem, for which the textbook solution uses hash tables. Then, I will talk about detecting the connectivity of graphs using distributed sketching, and my recent work showing the optimality of a well-known sketching algorithm (the AGM sketch). I will conclude the talk with discussion on future directions and my other work in theoretical computer science. Bio: Huacheng Yu is an associate research scholar in the Department of Computer Science at Princeton University. His research interests include data structures and streaming algorithms, and other directions in theoretical computer science such as communication complexity and graph algorithms. Prior to Princeton, Huacheng was a postdoctoral researcher at Harvard University hosted by Jelani Nelson and Madhu Sudan. He received his Ph.D. from Stanford University (advised by Ryan Williams and Omer Reingold) and B.Eng from Tsinghua University, both in Computer Science.

1 0