- talks - lists.cs.princeton.edu

Reminder: Colloquium Speaker Dylan Hadfield-Menell, tomorrow at 12:30pm
by Emily C. Lawrence 23 Feb '20

23 Feb '20

CS Colloquium Speaker Dylan Hadfield- Menell , University of California, Berkeley Monday, February 24 - 12:30pm Computer Science - Room 105 Host: Tom Griffiths [ https://www.cs.princeton.edu/events/25920 | https://www.cs.princeton.edu/events/25920 ] Title: The Value Alignment Problem in Artificial Intelligence Much of our success in artificial intelligence stems from the adoption of a simple paradigm: specify an objective or goal, and then use optimization algorithms to identify a behavior (or predictor) that optimally achieves this goal. This has been true since the early days of AI (e.g., search algorithms such as A* that aim to find the optimal path to a goal state), and this paradigm is common to AI, statistics, control theory, operations research, and economics. Loosely speaking, the field has evaluated the intelligence of an AI system by how efficiently and effectively it optimizes for its objective. This talk will provide an overview of my thesis work, which proposes and explores the consequences of a simple, but consequential, shift in perspective: we should measure the intelligence of an AI system by its ability to optimize for our objectives. In an ideal world, these measurements would be the same -- all we have to do is write down the correct objective! This is easier said than done: misalignment between the behavior a system designer actually wants and the behavior incentivized by the reward or loss functions they specify is routine, it is commonly observed in a wide variety of practical applications, and fundamental, as a consequence of limited human cognitive capacity. This talk will build up a formal model of this value alignment problem as a cooperative human-robot interaction: an assistance game of partial information between a human principal and an autonomous agent. It will begin with a discussion of a simple instantiation of this game where the human designer takes one action, write down a proxy objective, and the robot attempts to optimize for the true objective by treating the observed proxy as evidence about the intended goal. Next, I will generalize this model to introduce Cooperative Inverse Reinforcement Learning, a general and formal model of this assistance game, and discuss the design of efficient algorithms to solve it. The talk will conclude with a discussion of directions for further research including applications to content recommendation and home robotics, the development of reliable and robust design environments for AI objectives, and the theoretical study of AI regulation by society as a value alignment problem with multiple human principals. Bio: Dylan is a final year Ph.D. student at UC Berkeley, advised by Anca Dragan, Pieter Abbeel , and Stuart Russell. His research focuses on the value alignment problem in artificial intelligence. His goal is to design algorithms that learn about and pursue the intended goal of their users, designers, and society in general. His recent work has focused on algorithms for human-robot interaction with unknown preferences and reliability engineering for learning systems.

1 0

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

20 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

CS and CITP Colloquium Speaker Aleksandra Korolova, Tues Feb 25 at 12:30pm
by Emily C. Lawrence 20 Feb '20

20 Feb '20

CS and CITP Colloquium Speaker Aleksandra Korolova , University of Southern California Tuesday, February 25 - 12:30pm Computer Science - Room 105 Host: Jonathan Mayer [ https://www.cs.princeton.edu/events/25931 | https://www.cs.princeton.edu/events/25931 ] Title: Aligning the web ecosystem with societal values via measurements, definitions and algorithms The web ecosystem relies on data- and algorithms- driven innovation and monetization. In this talk, I will explore the challenges and opportunities for making these practices aligned with societal values such as privacy and fairness. I will first describe our work in applying differential privacy for development of data-driven product features without requiring raw data collection from users. I will then present our work on systematically identifying new societal concerns posed by current implementations of targeted advertising systems such as: privacy violations, manipulation of the vulnerable, discrimination, and creation of informational filter bubbles. I will conclude by discussing how such measurements, combined with definitions and algorithms, particularly from the fields of differential privacy and algorithmic fairness, can provide a language for reasoning about trade-offs between utility, privacy, and fairness, and give solution approaches to both computer scientists and policy makers. Bio: Aleksandra Korolova is a WiSE Gabilan Assistant Professor of Computer Science at USC, where she researches algorithms and technologies that enable data-driven innovations while preserving privacy and fairness. Prior to joining USC, Aleksandra was a research scientist at Google. Aleksandra received her PhD in Computer Science from Stanford University. Her PhD thesis, which focused on protecting privacy when mining and sharing user data, has been recognized by the Arthur L. Samuel Thesis Award 2011-2012, for the best PhD thesis in the Computer Science Department at Stanford. Aleksandra is also a co-winner of the 2011 PET Award for outstanding research in privacy enhancing technologies for exposing privacy violations of microtargeted advertising and a runner-up for the 2015 PET Award for RAPPOR, the first commercial deployment of differential privacy. Aleksandra's most recent work, on discrimination in ad delivery, has received CSCW Honorable Mention Award and Recognition of Contribution to Diversity and Inclusion, was cited in Facebook's Civil Rights Audit Report, and invited for a briefing for Members of the House Financial Services Committee. This talk is being co-sponsored by CITP and the Department of Computer Science.

1 0

Colloquium Speaker Dylan Hadfield-Menell, Monday Feb 24 at 12:30pm
by Emily C. Lawrence 19 Feb '20

19 Feb '20

CS Colloquium Speaker Dylan Hadfield- Menell , University of California, Berkeley Monday, February 24 - 12:30pm Computer Science - Room 105 Host: Tom Griffiths [ https://www.cs.princeton.edu/events/25920 | https://www.cs.princeton.edu/events/25920 ] Title: The Value Alignment Problem in Artificial Intelligence Much of our success in artificial intelligence stems from the adoption of a simple paradigm: specify an objective or goal, and then use optimization algorithms to identify a behavior (or predictor) that optimally achieves this goal. This has been true since the early days of AI (e.g., search algorithms such as A* that aim to find the optimal path to a goal state), and this paradigm is common to AI, statistics, control theory, operations research, and economics. Loosely speaking, the field has evaluated the intelligence of an AI system by how efficiently and effectively it optimizes for its objective. This talk will provide an overview of my thesis work, which proposes and explores the consequences of a simple, but consequential, shift in perspective: we should measure the intelligence of an AI system by its ability to optimize for our objectives. In an ideal world, these measurements would be the same -- all we have to do is write down the correct objective! This is easier said than done: misalignment between the behavior a system designer actually wants and the behavior incentivized by the reward or loss functions they specify is routine, it is commonly observed in a wide variety of practical applications, and fundamental, as a consequence of limited human cognitive capacity. This talk will build up a formal model of this value alignment problem as a cooperative human-robot interaction: an assistance game of partial information between a human principal and an autonomous agent. It will begin with a discussion of a simple instantiation of this game where the human designer takes one action, write down a proxy objective, and the robot attempts to optimize for the true objective by treating the observed proxy as evidence about the intended goal. Next, I will generalize this model to introduce Cooperative Inverse Reinforcement Learning, a general and formal model of this assistance game, and discuss the design of efficient algorithms to solve it. The talk will conclude with a discussion of directions for further research including applications to content recommendation and home robotics, the development of reliable and robust design environments for AI objectives, and the theoretical study of AI regulation by society as a value alignment problem with multiple human principals. Bio: Dylan is a final year Ph.D. student at UC Berkeley, advised by Anca Dragan, Pieter Abbeel , and Stuart Russell. His research focuses on the value alignment problem in artificial intelligence. His goal is to design algorithms that learn about and pursue the intended goal of their users, designers, and society in general. His recent work has focused on algorithms for human-robot interaction with unknown preferences and reliability engineering for learning systems.

1 0

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