February 2020 - talks - lists.cs.princeton.edu

Colloquium Speaker Zhihao Jia, TODAY at 12:30pm
by Emily C. Lawrence 19 Feb '20

19 Feb '20

CS Colloquium Speaker Zhihao Jia, Stanford University Wednesday, February 19 - 12:30pm Computer Science - Room 105 Host: Kai Li [ https://www.cs.princeton.edu/events/25916 | https://www.cs.princeton.edu/events/25916 ] Title: Automated Discovery of Machine Learning Optimizations As an increasingly important workload, machine learning (ML) applications require different performance optimization techniques from traditional runtimes and compilers. In particular, to accelerate ML applications, it is generally necessary to perform ML computations on heterogeneous hardware and parallelize computations using multiple data dimensions, neither of which is even expressible in traditional compilers and runtimes. In this talk, I will describe my work on automated discovery of performance optimizations to accelerate ML computations. TASO, the Tensor Algebra SuperOptimizer , optimizes the computation graphs of deep neural networks (DNNs) by automatically generating potential graph optimizations and formally verifying their correctness. TASO outperforms rule-based graph optimizers in existing ML systems (e.g., TensorFlow , TensorRT , and TVM) by up to 3x by automatically discovering novel graph optimizations, while also requiring significantly less human effort. FlexFlow is a system for accelerating distributed DNN training. FlexFlow identifies parallelization dimensions not considered in existing ML systems (e.g., TensorFlow and PyTorch ) and automatically discovers fast parallelization strategies for a specific parallel machine. Companies and national labs are using FlexFlow to train production ML models that do not scale well in current ML systems, achieving over 10x performance improvement. I will also outline future research directions for further automating ML systems, such as codesigning ML models, software systems, and hardware backends for end-to-end ML deployment. Bio: Zhihao Jia is a Ph.D. candidate in the Computer Science department at Stanford University working with Alex Aiken and Matei Zaharia . His research interests lie in the intersection of computer systems and machine learning, with a focus on building efficient, scalable, and high-performance systems for ML computations.

1 0

Reminder: Colloquium Speaker Aviv Tamar, tomorrow at 12:30pm
by Emily C. Lawrence 19 Feb '20

19 Feb '20

CS Colloquium Speaker Aviv Tamar, Technion - Israel Institute for Technology Thursday, February 20 - 12:30pm Computer Science - Room 105 Host: Elad Hazan [ https://www.cs.princeton.edu/events/25919 | https://www.cs.princeton.edu/events/25919 ] Title: Robot Learning and Planning How can we build autonomous robots that operate in unstructured and dynamic environments such as homes or hospitals? This problem has been investigated under several disciplines, including planning (motion planning, task planning, etc.), and reinforcement learning. While both of these fields have witnessed tremendous progress, each have fundamental drawbacks: planning approaches require substantial manual engineering in mapping perception to a formal planning problem, while RL, which can operate directly on raw percepts, is data hungry, cannot generalize to new tasks, and is ‘black box’ in nature. In this talk, we present several studies that aim to mitigate these shortcomings by combining ideas from both planning and learning. We start by introducing value iteration networks, a type of differentiable planner that can be used within model-free RL to obtain better generalization. Next, we consider a practical robotic assembly problem, and show that motion planning, based on readily available CAD data, can be combined with RL to quickly learn policies for assembling tight fitting objects. We conclude with our recent work on learning to imagine goal-directed visual plans. Motivated by humans’ remarkable capability to predict and plan complex manipulations of objects, and recent advances such as GANs in imagining images, we present Visual Plan Imagination (VPI) — a new computational problem that combines image imagination and planning. In VPI, given off-policy image data from a dynamical system, the task is to ‘imagine’ image sequences that transition the system from start to goal. Key to our method is Causal InfoGAN , a deep generative model that can learn features that are compatible with strong planning algorithms. We demonstrate our approach on learning to imagine and execute robotic rope manipulation from image data. Bio: Aviv Tamar is an Assistant Professor at the Department of Electrical Engineering at Technion - Israel Institute for Technology. Previously, he was a post-doc at UC Berkeley with Prof. Pieter Abbeel , and prior to that, he completed his PhD with Prof. Shie Mannor at Technion . Aviv's research focuses on reinforcement learning, representation learning, and robotics. His work has been recognized by a NeurIPS Best Paper award, a Google Faculty Award, and the Alon fellowship for young researchers.

1 0

Reminder: Colloquium Speaker Zhihao Jia, tomorrow at 12:30pm
by Emily C. Lawrence 18 Feb '20

18 Feb '20

CS Colloquium Speaker Zhihao Jia, Stanford University Wednesday, February 19 - 12:30pm Computer Science - Room 105 Host: Kai Li [ https://www.cs.princeton.edu/events/25916 | https://www.cs.princeton.edu/events/25916 ] Title: Automated Discovery of Machine Learning Optimizations As an increasingly important workload, machine learning (ML) applications require different performance optimization techniques from traditional runtimes and compilers. In particular, to accelerate ML applications, it is generally necessary to perform ML computations on heterogeneous hardware and parallelize computations using multiple data dimensions, neither of which is even expressible in traditional compilers and runtimes. In this talk, I will describe my work on automated discovery of performance optimizations to accelerate ML computations. TASO, the Tensor Algebra SuperOptimizer , optimizes the computation graphs of deep neural networks (DNNs) by automatically generating potential graph optimizations and formally verifying their correctness. TASO outperforms rule-based graph optimizers in existing ML systems (e.g., TensorFlow , TensorRT , and TVM) by up to 3x by automatically discovering novel graph optimizations, while also requiring significantly less human effort. FlexFlow is a system for accelerating distributed DNN training. FlexFlow identifies parallelization dimensions not considered in existing ML systems (e.g., TensorFlow and PyTorch ) and automatically discovers fast parallelization strategies for a specific parallel machine. Companies and national labs are using FlexFlow to train production ML models that do not scale well in current ML systems, achieving over 10x performance improvement. I will also outline future research directions for further automating ML systems, such as codesigning ML models, software systems, and hardware backends for end-to-end ML deployment. Bio: Zhihao Jia is a Ph.D. candidate in the Computer Science department at Stanford University working with Alex Aiken and Matei Zaharia . His research interests lie in the intersection of computer systems and machine learning, with a focus on building efficient, scalable, and high-performance systems for ML computations.

1 0

Colloquium Speaker Sara Achour - TODAY at 12:30pm
by Emily C. Lawrence 18 Feb '20

18 Feb '20

CS Colloquium Speaker Sara Achour, Massachusetts Institute of Technology Tuesday, February 18 - 12:30pm Computer Science - Room 105 Host: Jennifer Rexford [ https://www.cs.princeton.edu/events/25917 | https://www.cs.princeton.edu/events/25917 ] Title: New Compilation Techniques for Reconfigurable Analog Devices Reconfigurable analog devices are a powerful new computing substrate especially appropriate for executing dynamical systems in an energy efficient manner. These devices leverage the physical behavior of transistors to directly implement computation. Under this paradigm, voltages and currents within the device implement continuously evolving variables in the computation. In this talk, I discuss compilation techniques for automatically configuring such devices to execute dynamical systems. I present Legno , the first compilation system that automatically targets a real reconfigurable analog device of this class. Legno synthesizes analog circuits from parametric and specialized analog blocks and accounts for analog noise, quantization error, operating range limitations, and manufacturing variations within the device. I evaluate Legno on applications from the biology, physics, and controls domains. The results demonstrate that these applications execute with acceptable error while consuming microjoules of energy. Bio: Sara Achour is a PhD candidate at the Computer Science and Artificial Intelligence Laboratory at Massachusetts Institute of Technology (CSAIL MIT) and a NSF Fellowship recipient. Her research focuses on new techniques and tools, specifically new programming languages, compilers, and runtime systems, that enable end users to easily develop computations that exploit the potential of emerging nontraditional computing platforms.

1 0

Sumegha Garg will present her Pre FPO "Implications of Space-Bounded Computation" on Friday, February 21, 2020 at 10am in CS 402.
by Nicki Mahler 17 Feb '20

17 Feb '20

Sumegha Garg will present her Pre FPO "Implications of Space-Bounded Computation" on Friday, February 21, 2020 at 10am in CS 402. The members of her committee are as follows: Mark Braverman (advisor); Readers: Mark Braverman and Ran Raz; Examiners: Gillat Kol, Matt Weinberg, and Mark Zhandry. Everyone is invited to her talk. Please see below for the talk abstract. Abstract : The field of complexity theory studies the inherent limits of a computational model for performing certain computational tasks. Most natural is to study restricted computational models that are short of a computational resource, for example, time or memory. While characterizing the minimum time required for a computational task has received more attention, understanding the memory requirements is as fundamental and fascinating, owing to the surprising and non-intuitive results that the field has seen. In my thesis, I focus on understanding the limits of space-bounded computational models and the implications for various algorithmic problems. Implications of bounded space for learning algorithms : [Sha14, SVW16] started the study of feasibility of online learning under memory constraints. In a breakthrough result, [Raz16] showed that learning a n-bit vector from random linear equations requires either Ω(n^2) space or 2^{Ω(n)} samples. My work has focused on extending these memory-sample tradeoffs to a larger class of learning problems and to even when the learner is allowed a second pass over the stream of samples (through an extractor-based approach). Implications of bounded space for randomized algorithms : It is largely unknown if randomization gives space-bounded algorithms any advantage over deterministic computation. The current best hope of the community is to derandomize randomized log-space computation with one-sided error, that is, prove RL = L. In a recent work, a step towards answering this question, we improve upon the previous constructions of hitting sets, which are tools for derandomization. Implications of bounded space for streaming games : In a recent work, we show the impossibility of winning a simple streaming game under memory constraints. In this game, Alice and Bob take turns saying numbers belonging to the set {1, 2, ..., 2N }. A player loses if they repeat a number that has already been said. Bob, who goes second, has a very simple (and memoryless) strategy to avoid losing. We prove that Alice, however, needs at least Ω(N) memory to win.

1 0

Niranjani Prasad will present her Pre FPO "Methods for Reinforcement Learning in Clinical Decision Support" on Friday, February 21, 2020 at 1pm in CS 402
by Nicki Mahler 17 Feb '20

17 Feb '20

Niranjani Prasad will present her Pre FPO "Methods for Reinforcement Learning in Clinical Decision Support" on Friday, February 21, 2020 at 1pm in CS 402. The members of her committee are as follows: Barbara Engelhardt (advisor); Readers: Ryan Adams, Finale Doshi-Velez (Harvard); Examiners: Sebastian Seung, Mengdi Wang (ORFE), and Barbara Engelhardt Everyone is invited to attend her talk. The talk abstract follows below: The management of routine interventions such as mechanical ventilation and sedation, or the ordering of blood tests, constitutes a major part of the care of critically ill patients. Timely and proportionate interventions are crucial to improving patient outcomes and reducing costs, but these procedures are often poorly understood, particularly for heterogeneous patient populations, and clinical opinion on best protocols vary. In my thesis, I focus on the development of a clinician-in-loop decision support system for weaning patients from mechanical ventilation, that leverages available information in the data-intensive ICU setting to predict time to extubation readiness and provide personalized recommendations for sedation and ventilator support. To this end, I employ off-policy reinforcement learning algorithms to learn an optimal sequence of treatment actions from suboptimal historical ICU data. I model patient admissions as Markov decision processes (MDPs), developing tailored representations of patient state, action space and reward function, and learn treatment policies using Fitted Q-iteration. I demonstrate that this framework shows promise in recommending protocols with positive clinical outcomes, in terms of managing patient stability and reintubation rates, when assessed against current practice. The second part of my thesis is directed towards the problem of effective reward design when applying reinforcement learning to clinical decision-making tasks. A key impediment to reinforcement learning in practice is in distilling multiple (and often conflicting) clinical imperatives into a single scalar feedback signal. To tackle this, I develop methods based on Pareto optimality to solve MDPs given a vector-valued reward function. I illustrate these methods in the context of optimizing the ordering of laboratory tests for patients in critical care. Extending this further, I introduce a mechanism for using available clinical data to restrict the candidate set of convex combinations of elements in the reward function, to those that yield a scalar reward which (a) reflects what we implicitly know from the data about reasonable behaviour for a task, and (b) allows for robust off-policy evaluation. Together, these help ensure that we learn policies that we trust to be implemented in high-risk settings.

1 0

Reminder: Colloquium Speaker Sara Achour - tomorrow at 12:30pm
by Emily C. Lawrence 17 Feb '20

17 Feb '20

CS Colloquium Speaker Sara Achour, Massachusetts Institute of Technology Tuesday, February 18 - 12:30pm Computer Science - Room 105 Host: Jennifer Rexford [ https://www.cs.princeton.edu/events/25917 | https://www.cs.princeton.edu/events/25917 ] Title: New Compilation Techniques for Reconfigurable Analog Devices Reconfigurable analog devices are a powerful new computing substrate especially appropriate for executing dynamical systems in an energy efficient manner. These devices leverage the physical behavior of transistors to directly implement computation. Under this paradigm, voltages and currents within the device implement continuously evolving variables in the computation. In this talk, I discuss compilation techniques for automatically configuring such devices to execute dynamical systems. I present Legno , the first compilation system that automatically targets a real reconfigurable analog device of this class. Legno synthesizes analog circuits from parametric and specialized analog blocks and accounts for analog noise, quantization error, operating range limitations, and manufacturing variations within the device. I evaluate Legno on applications from the biology, physics, and controls domains. The results demonstrate that these applications execute with acceptable error while consuming microjoules of energy. Bio: Sara Achour is a PhD candidate at the Computer Science and Artificial Intelligence Laboratory at Massachusetts Institute of Technology (CSAIL MIT) and a NSF Fellowship recipient. Her research focuses on new techniques and tools, specifically new programming languages, compilers, and runtime systems, that enable end users to easily develop computations that exploit the potential of emerging nontraditional computing platforms.

1 0

Colloquium Speaker Aviv Tamar, Thurs Feb 20 at 12:30pm
by Emily C. Lawrence 14 Feb '20

14 Feb '20

CS Colloquium Speaker Aviv Tamar, Technion - Israel Institute for Technology Thursday, February 20 - 12:30pm Computer Science - Room 105 Host: Elad Hazan [ https://www.cs.princeton.edu/events/25919 | https://www.cs.princeton.edu/events/25919 ] Title: Robot Learning and Planning How can we build autonomous robots that operate in unstructured and dynamic environments such as homes or hospitals? This problem has been investigated under several disciplines, including planning (motion planning, task planning, etc.), and reinforcement learning. While both of these fields have witnessed tremendous progress, each have fundamental drawbacks: planning approaches require substantial manual engineering in mapping perception to a formal planning problem, while RL, which can operate directly on raw percepts, is data hungry, cannot generalize to new tasks, and is ‘black box’ in nature. In this talk, we present several studies that aim to mitigate these shortcomings by combining ideas from both planning and learning. We start by introducing value iteration networks, a type of differentiable planner that can be used within model-free RL to obtain better generalization. Next, we consider a practical robotic assembly problem, and show that motion planning, based on readily available CAD data, can be combined with RL to quickly learn policies for assembling tight fitting objects. We conclude with our recent work on learning to imagine goal-directed visual plans. Motivated by humans’ remarkable capability to predict and plan complex manipulations of objects, and recent advances such as GANs in imagining images, we present Visual Plan Imagination (VPI) — a new computational problem that combines image imagination and planning. In VPI, given off-policy image data from a dynamical system, the task is to ‘imagine’ image sequences that transition the system from start to goal. Key to our method is Causal InfoGAN , a deep generative model that can learn features that are compatible with strong planning algorithms. We demonstrate our approach on learning to imagine and execute robotic rope manipulation from image data. Bio: Aviv Tamar is an Assistant Professor at the Department of Electrical Engineering at Technion - Israel Institute for Technology. Previously, he was a post-doc at UC Berkeley with Prof. Pieter Abbeel , and prior to that, he completed his PhD with Prof. Shie Mannor at Technion . Aviv's research focuses on reinforcement learning, representation learning, and robotics. His work has been recognized by a NeurIPS Best Paper award, a Google Faculty Award, and the Alon fellowship for young researchers.

1 0

Colloquium Speaker Zhihao Jia, Wed Feb 19 at 12:30pm
by Emily C. Lawrence 13 Feb '20

13 Feb '20

CS Colloquium Speaker Zhihao Jia, Stanford University Wednesday, February 19 - 12:30pm Computer Science - Room 105 Host: Kai Li [ https://www.cs.princeton.edu/events/25916 | https://www.cs.princeton.edu/events/25916 ] Title: Automated Discovery of Machine Learning Optimizations As an increasingly important workload, machine learning (ML) applications require different performance optimization techniques from traditional runtimes and compilers. In particular, to accelerate ML applications, it is generally necessary to perform ML computations on heterogeneous hardware and parallelize computations using multiple data dimensions, neither of which is even expressible in traditional compilers and runtimes. In this talk, I will describe my work on automated discovery of performance optimizations to accelerate ML computations. TASO, the Tensor Algebra SuperOptimizer , optimizes the computation graphs of deep neural networks (DNNs) by automatically generating potential graph optimizations and formally verifying their correctness. TASO outperforms rule-based graph optimizers in existing ML systems (e.g., TensorFlow , TensorRT , and TVM) by up to 3x by automatically discovering novel graph optimizations, while also requiring significantly less human effort. FlexFlow is a system for accelerating distributed DNN training. FlexFlow identifies parallelization dimensions not considered in existing ML systems (e.g., TensorFlow and PyTorch ) and automatically discovers fast parallelization strategies for a specific parallel machine. Companies and national labs are using FlexFlow to train production ML models that do not scale well in current ML systems, achieving over 10x performance improvement. I will also outline future research directions for further automating ML systems, such as codesigning ML models, software systems, and hardware backends for end-to-end ML deployment. Bio: Zhihao Jia is a Ph.D. candidate in the Computer Science department at Stanford University working with Alex Aiken and Matei Zaharia . His research interests lie in the intersection of computer systems and machine learning, with a focus on building efficient, scalable, and high-performance systems for ML computations.

1 0

Jennifer Gossels will present her FPO, "Joint Optimization for Robust Network Design and Operation" on Wednesday, 2/19/2020 at 9:30am in CS 402
by Nicki Mahler 12 Feb '20

12 Feb '20

Jennifer Gossels will present her FPO, "Joint Optimization for Robust Network Design and Operation" on Wednesday, 2/19/2020 at 9:30am in CS 402. The members of her committee are as follows: Jennifer Rexford (adviser), Readers: Kai Li and Gagan Choudhury (AT &T Labs); Examiners: Michael Freedman, Wyatt Lloyd, and Jennifer Rexford A copy of her thesis, is available upon request. Please email ngotsis(a)cs.princeton if you would like a copy of the thesis. Everyone is invited to attend her talk. The talk abstract follows below: The Internet is an essential part of modern life, and Internet Service Provider (ISP) backbone networks are integral to our Internet experience. Therefore, ISPs must build networks that limit congestion, even when some equipment fails. This network design problem is complicated, because an optimal network design must consider the eventual runtime configuration. An ISP makes network design decisions, such as purchasing and placing equipment, on long timescales (months or years) and network operation decisions, such as routing packets, on short timescales (seconds). Design and operation interact such that the ISP must solve the network operation problem as a sub-problem of network design, rendering the network design problem difficult to formulate and computationally complex. Today ISPs resort to a variety of simplifications; they fail to take advantage of the reconfigurability offered by modern optical equipment or the opportunity to mix more- and less-powerful switches throughout their networks. In this dissertation, we show how ISPs can incorporate each of these factors into their network design and operation models to produce less expensive networks without compromising robustness. In Chapter 2, we explain how reconfigurable optical switches fundamentally change the network design and operation problems by shifting the boundary between what is fixed at design time and what is reconfigured at runtime. Then, we present an optimal formulation for this new problem and heuristics to help our solution scale. In Chapter 3, we describe a failure recovery protocol that allows ISPs to realize many of the benefits of outfitting their networks with a homogeneous collection of powerful, “smart” switches, while instead using a combination of these expensive boxes and less expensive, “dumb” switches. We make three contributions in each chapter. First, we formulate the network design optimization by extending the multicommodity flow framework to leverage colorless and directionless Reconfigurable Optical Add/Drop Multiplexers (Chapter 2) or heterogeneous nodes (Chapter 3). Second, we devise heuristics to scale our designs to larger topologies. Finally, we evaluate our ideas on a realistic backbone topology and traffic demands.

1 0