Quick links

CS Department Colloquium Series

Robot Learning and Planning

Date and Time
Thursday, February 20, 2020 - 12:30pm to 1:30pm
Location
Computer Science Small Auditorium (Room 105)
Type
CS Department Colloquium Series
Host
Elad Hazan

Aviv Tamar
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.

*Please note, this event is only open to the Princeton University community.

Lunch for talk attendees will be available at 12:00pm. 
To request accommodations for a disability, please contact Emily Lawrence, emilyl@cs.princeton.edu, 609-258-4624 at least one week prior to the event.

Building Machines that Discover Generalizable, Interpretable Knowledge

Date and Time
Tuesday, February 11, 2020 - 12:30pm to 1:30pm
Location
Computer Science Small Auditorium (Room 105)
Type
CS Department Colloquium Series
Host
Tom Griffiths

Kevin Ellis
Humans can learn to solve an endless range of problems: building, drawing, designing, coding, and cooking, to name a few, and need relatively modest amounts of experience to acquire any one new individual skill. Machines which can similarly master a diverse span of problems are surely far off.

Here, however, I will argue that program induction--an emerging AI technique--will play a role in building this more human-like AI. Program induction systems represent knowledge as programs, and learn by synthesizing code. Across three case studies in vision, natural language, and learning-to-learn, this talk will present program induction systems that take a step toward machines that can: acquire new knowledge from modest amounts of experience; strongly generalize that knowledge to extrapolate beyond their training; learn to represent their knowledge in an interpretable format; and are applicable to a broad spread of problems, from drawing pictures to discovering equations. Driving these developments is a new neuro-symbolic algorithm for Bayesian program synthesis. This algorithm integrates maturing program synthesis technologies with several complementary AI traditions (symbolic, probabilistic, and neural).

Building a human-like machine learner is a very distant, long-term goal for the field. In the near-term program induction comes with a roadmap of practical problems to push, such as language learning, scene understanding, and programming-by-examples, which this talk explores. But it's worth keeping these long-term goals in mind as well.

Bio: Kevin Ellis works across artificial intelligence, program synthesis, and machine learning. He develops learning algorithms that teach machines to write code, and applies these algorithms to problems in artificial intelligence. His work has appeared in machine learning venues (NeurIPS, ICLR, IJCAI) and cognitive science venues (CogSci, TOPICS). He has collaborated with researchers at Harvard, Brown, McGill, Siemens, and MIT, where he is a final-year graduate student advised by Josh Tenenbaum and Armando Solar-Lezama.

*Please note, this event is only open to the Princeton University community.

Lunch for talk attendees will be available at 12:00pm. 
To request accommodations for a disability, please contact Emily Lawrence, emilyl@cs.princeton.edu, 609-258-4624 at least one week prior to the event.

New Compilation Techniques for Reconfigurable Analog Devices

Date and Time
Tuesday, February 18, 2020 - 12:30pm to 1:30pm
Location
Computer Science Small Auditorium (Room 105)
Type
CS Department Colloquium Series
Host
Jennifer Rexford

Sara Achour
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.

*Please note, this event is only open to the Princeton University community.

Lunch for talk attendees will be available at 12:00pm. 
To request accommodations for a disability, please contact Emily Lawrence, emilyl@cs.princeton.edu, 609-258-4624 at least one week prior to the event.

Automated Discovery of Machine Learning Optimizations

Date and Time
Wednesday, February 19, 2020 - 12:30pm to 1:30pm
Location
Computer Science Small Auditorium (Room 105)
Type
CS Department Colloquium Series
Host
Kai Li

Zhihao Jia
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.

*Please note, this event is only open to the Princeton University community.

Lunch for talk attendees will be available at 12:00pm. 
To request accommodations for a disability, please contact Emily Lawrence, emilyl@cs.princeton.edu, 609-258-4624 at least one week prior to the event.

Project Debater - How persuasive can a computer be?

Date and Time
Friday, February 7, 2020 - 12:30pm to 1:30pm
Location
Computer Science Small Auditorium (Room 105)
Type
CS Department Colloquium Series
Host
Yoram Singer and Karthik Narasimhan

Project Debater is the first AI system that can meaningfully debate a human opponent. The system, an IBM Grand Challenge, is designed to build coherent, convincing speeches on its own, as well as provide rebuttals to the opponent's main arguments. In February 2019, Project Debater competed against Harish Natarajan, who holds the world record for most debate victories, in an event held in San Francisco that was broadcasted live world-wide. In this talk I will tell the story of Project Debater, from conception to a climatic final event, describe its underlying technology, and discuss how it can be leveraged for advancing decision making and critical thinking.

Bio:​ Noam Slonim is a Distinguished Engineer at IBM Research AI. He received his doctorate from the Interdisciplinary Center for Neural Computation at the Hebrew University and held a post-doc position at the Genomics Institute at Princeton University. During his PhD, Noam received the best paper award in UAI and ECIR, and the best presentation award at SIGIR. Noam joined the IBM Haifa Research Lab in 2007, and in 2011 he proposed to develop Project Debater. He has been serving as the Principal Investigator of the project since then. Noam published around 60 peer reviewed articles, focusing on the last few years on advancing the emerging field of Computational Argumentation. Finally, Noam used to have a secondary career as a TV script writer. Coincidentally, or not, in a sitcom he co-created back in 1998, the last episode was focused on competitive debates.

Lunch for talk attendees will be available at 12:00pm. 
To request accommodations for a disability, please contact Emily Lawrence, emilyl@cs.princeton.edu, 609-258-4624 at least one week prior to the event.

Learning to predict the future in healthcare

Date and Time
Monday, February 10, 2020 - 12:30pm to 1:30pm
Location
Computer Science Small Auditorium (Room 105)
Type
CS Department Colloquium Series
Host
Elad Hazan

Kira Radinsky
Our world faces increasingly complex challenges: we destabilized the climate, haven’t beaten all diseases, and haven’t spread the values of democracy and freedom to large parts of the globe, where violence and riots reign supreme. The world must be fixed in our generation - everyone would agree. But in order to take action, build a plan, we need to see the complete picture, and empower decision makers with tools to make those changes. This decade, we have finally reached a critical amount of data to facilitate the creation of such tools.

My work is inspired by Mark Twain’s quote, who once said: “The past does not repeat itself, but it rhymes." Although future events have unique circumstances, they typically follow familiar past patterns. Over the past few years, I devoted my life to development of prediction techniques. My system inferred that Cholera outbreaks in land-locked areas are more likely to occur following storms, especially when preceded by a long drought. Another inference is that genocide events tend to occur following events where local opinion makers describe minority groups as pests. These types of patterns are composed of several abstractions, over variable-term temporal extents and selected from a large number of possible causes.  The algorithms I developed deal with the complexity of discovering such patterns.

Large-scale digital histories, social and real-time media, and human web behavior are harvested and augmented with human knowledge mined from the web to afford real-time estimations of likelihoods of future events. Most recently, these algorithms have accurately predicted the first Cholera outbreak reported in Cuba in fifty years. These types of actionable predictions, that enable preventative measures, have drawn the attention of a UN genocide-prevention organization and the Gates foundations and illustrate the vast potential for real impact on the state of humanity.

In the last few years I have been focusing on applying similar techniques for the healthcare and Pharma, leveraging large amount of data obtained from both medical records, EMR and other medical research results data in a quest to create an AI system for automated medical research and breakthroughs.

Bio: Dr. Kira Radinsky is the chairman and CTO of Diagnostic Robotics, where the most advanced technologies in the field of artificial intelligence are harnessed to make healthcare better, cheaper, and more widely available. Dr. Radinsky has founded SalesPredict, acquired by eBay in 2016 and served as eBay Chief Scientist (IL). She gained international recognition for her work at the Technion and Microsoft Research for developing predictive algorithms that recognized the early warning signs of globally impactful events, as disease epidemics and political unrests. In 2013, she was named one of MIT Technology Review’s 35 Young Innovators Under 35 and in 2015 Forbes included her as "30 Under 30 Rising Stars in Enterprise Tech”. She is a frequent presenter at global tech and industry conferences, including TEDx, Wired, Strata Data Science, Techcrunch and publishes in HBR. Radinsky also serves as a board member in Israel Security Authorities and technology board of HSBC bank. She also holds a visiting professor position at the Technion  focusing on the application of predictive data mining in medicine.

Lunch for talk attendees will be available at 12:00pm. 
To request accommodations for a disability, please contact Emily Lawrence, emilyl@cs.princeton.edu, 609-258-4624 at least one week prior to the event.

New Challenges in Question Answering: Natural Questions and Going Beyond Word Matching

Date and Time
Tuesday, December 3, 2019 - 1:30pm to 2:50pm
Location
Computer Science Large Auditorium (Room 104)
Type
CS Department Colloquium Series
Speaker
Host
Karthik Narasimhan

Recently, learned deep models have surpassed human performance on a number of question answering benchmarks such as SQuAD. However, these models resort to simple word matching and answer typing heuristics, and they are easily fooled. In this talk, I will present two different lines of work that aim to take us beyond the current status quo in question answering, and push us towards more robust representations of semantics, pragmatics, and knowledge.

First, I will present Natural Questions (NQ), a new question answering benchmark from Google. NQ contains real user questions, which require an understanding of the questioner's intent. NQ also requires systems to read entire Wikipedia pages to decide whether they fully answer a question. This is much harder than finding an answer given the knowledge that one is present. I will convince you that the question 'when was the last time a hurricane hit Massachusetts?' is under-specified with many reasonable answers, and I will tell you how we developed robust evaluation metrics to deal with this ambiguity.

In the second part of the talk I will present a complementary method of challenging today's question answering systems by removing access to evidence documents at inference time. Instead of building joint representations of questions and documents, we perform ahead of inference time reading and retrieve answers via fast maximum inner product search. I will show that this leads to large gains in accuracy and speed when finding answers in very large corpora. I will also show some preliminary results that show how our methods can be used to aggregate information from multiple diverse documents.

To request accommodations for a disability, please contact Emily Lawrence, emilyl@cs.princeton.edu, 609-258-4624 at least one week prior to the event.

Putting together the threads of conversational AI

Date and Time
Tuesday, November 26, 2019 - 1:30pm to 2:50pm
Location
Computer Science Large Auditorium (Room 104)
Type
CS Department Colloquium Series
Speaker
Host
Danqi Chen

Jason Weston
Maybe we don't have enough threads yet to knit together the whole, but let's try anyway! We present our view of what is necessary for conversational AI, and the pieces we have worked on so far to get there. In particular: software (ParlAI, a unified platform for dialogue research), various neural architectures for memory, reasoning, retrieval and generation, and interactive learning, tasks for employing personality (Persona-Chat), knowledge (Wizard of Wikipedia) and perception (Image-Chat), evaluation studies & techniques (Acute-Eval), and a recent competition (ConvAI2). We'll try not to be too pessimistic (!), but we'll also discuss how far we still have to go.

Bio: Jason Weston is a research scientist at Facebook, NY and a Visiting Research Professor at NYU. He earned his PhD in machine learning at Royal Holloway, University of London and at AT&T Research in Red Bank, NJ (advisors: Alex Gammerman, Volodya Vovk and Vladimir Vapnik) in 2000. From 2000 to 2001, he was a researcher at Biowulf technologies. From 2002 to 2003 he was a research scientist at the Max Planck Institute for Biological Cybernetics, Tuebingen, Germany. From 2003 to 2009 he was a research staff member at NEC Labs America, Princeton. From 2009 to 2014 he was a research scientist at Google, NY. His interests lie in statistical machine learning, with a focus on reasoning, memory, perception, interaction and communication. Jason has published over 100 papers, including best paper awards at ICML and ECML, and a Test of Time Award for his work "A Unified Architecture for Natural Language Processing: Deep Neural Networks with Multitask Learning", ICML 2008 (with Ronan Collobert). He was part of the YouTube team that won a National Academy of Television Arts & Sciences Emmy Award for Technology and Engineering for Personalized Recommendation Engines for Video Discovery. He was listed as the 16th most influential machine learning scholar at AMiner and one of the top 50 authors in Computer Science in Science.

To request accommodations for a disability, please contact Emily Lawrence, emilyl@cs.princeton.edu, 609-258-4624 at least one week prior to the event.

Leveraging Allelic Heterogeneity in Association Studies

Date and Time
Friday, November 22, 2019 - 12:30pm to 1:30pm
Location
Computer Science Small Auditorium (Room 105)
Type
CS Department Colloquium Series
Host
Ben Raphael

Eleazar Eskin
Over 100,000 variants have been implicated in human traits through genome-wide association studies (GWAS).  Virtually all reported GWAS identify associations by measuring the correlation between a single variant and a phenotype of interest. Recently, several studies reported that at many risk loci, there may exist multiple causal variants, a phenomenon referred to as allelic heterogeneity. Naturally, for a locus with multiple causal variants with small effect sizes, the standard association test is underpowered to detect the associations. Alternatively, an approach considering effects of multiple variants simultaneously may increase statistical power. Counterintuitively, most approaches that consider multiple variants in association studies find fewer associations than the single SNP association test.  This is due to the fact that most multiple variant methods assume a structure of allelic heterogeneity which is very different from what is observed in genetic studies.  In this work, we propose a new statistical method, Model-based Association test Reflecting causal Status (MARS), that tries to find an association between variants in risk loci and a phenotype, considering multiple variants at each locus. One of the main advantages of MARS is that it only requires the existing summary statistics to detect associated risk loci. Thus, MARS is applicable to any association study with summary statistics, even though individual level data is not available for the study. Utilizing extensive simulated data sets, we show that MARS increases the power of detecting true associated risk loci compared to previous approaches that consider multiple variants, while robustly controls the type I error. Applied to data of 44 tissues provided by the Genotype- Tissue Expression (GTEx) consortium, we show that MARS identifies more eGenes compared to previous approaches in most of the tissues; e.g. MARS identified 16% more eGenes than the ones reported by the GTEx consortium. Moreover, applied to Northern Finland Birth Cohort (NFBC) data, we demonstrate that MARS effectively identifies association loci with improved power (56% of more loci found by MARS) in GWAS studies compared to the standard association test.

Bio: Dr. Eleazar Eskin serves as the inaugural chair for the UCLA Department of Computational Medicine. Fascinated by the intersection of computer science and biology, Dr. Eskin is researching and developing computational methods for the analysis of genetic variations in human disease. There are millions of variants in the genome and identifying the variants involved in disease often requires tens of thousands of patient samples. In order to analyze these tremendously large datasets, Dr. Eskin and his team are solving challenging computational problems and developing new computational techniques. He received his PhD in computer science from Columbia University. A recipient of the Alfred P. Sloan Foundation Research Fellowship, Dr. Eskin’s work is supported by the National Science Foundation and the National Institutes of Health.

Lunch for talk attendees will be available at 12:00pm. 
To request accommodations for a disability, please contact Emily Lawrence, emilyl@cs.princeton.edu, 609-258-4624 at least one week prior to the event.

Network Level IoT Security

Date and Time
Friday, November 22, 2019 - 11:00am to 12:00pm
Location
Computer Science Small Auditorium (Room 105)
Type
CS Department Colloquium Series
Host
Jennifer Rexford

David Hay
Computer networks have undergone and continue to experience a major transformation, whereby billions of low-cost devices are being connected to the network to provide additional functionality and better user experience. Unlike traditional network devices, these devices, collectively known as the ``Internet of Things'' (IoT), typically have very limited computational, memory, and power resources. These IoT devices became a major security concerns, both due to human factors and to technical challenges in deploying security mechanisms on devices with low resources. The number and diversity of IoT devices creates a huge attack surface that is often exploited by attackers to launch large-scale attacks, sometimes exploiting well-known vulnerabilities.

This talk will highlight the security concerns of IoT devices from a networking perspective and explore how to secure IoT devices using whitelists, in which communication between a device and an endpoint is prohibited unless that endpoint appears in the corresponding whitelist.  Finally, we will discuss deployment options for such a solution (namely, within the internet gateway, as virtual network function within the ISP network, or a combination of the two). 

Bio: David Hay is an Associate Professor with the Rachel and Selim Benin School of Computer Science and Engineering, Hebrew University, Jerusalem, Israel. He received the B.A. (summa cum laude) and Ph.D. degrees in computer science from the Technion—Israel Institute of Technology, Haifa, Israel, in 2001 and 2007, respectively. In addition, he was with IBM Haifa Research Labs, Haifa, Israel; Cisco Systems, San Jose, CA, USA; the Electronic Department, Politecnico di Torino, Turin, Italy; and the Electrical Engineering Department with Columbia University, New York, NY, USA. In 2010, he co-founded (with Prof. Brembler-Barr) the DEEPNESS lab, focusing on deep packet inspection in next-generation network devices. He has served as a technical program committee member of numerous networking conferences, and since 2018 serves as en editor of ACM/IEEE Transactions on Networking. His research interests are in computer networks—in particular, network algorithmics, packet classification, deep packet inspection, network survivability and resilience, software-defined networking, network-function virtualization, and various aspects of network security. 

To request accommodations for a disability, please contact Emily Lawrence, emilyl@cs.princeton.edu, 609-258-4624 at least one week prior to the event.

Follow us: Facebook Twitter Linkedin