Forum for Artificial Intelligence
Archive
|
This website is the archive for past Forum for Artificial Intelligence talks. Please click this link to navigate to the list of current talks. FAI meets every other week (or so) to discuss scientific, philosophical, and cultural issues in artificial intelligence. Both technical research topics and broader inter-disciplinary aspects of AI are covered, and all are welcome to attend! If you would like to be added to the FAI mailing list, subscribe here. If you have any questions or comments, please send email to Catherine Andersson. |
[ Upcoming talks ]
Monday, October 5, 2015, 11:00AM
|
Towards Robotics Algorithms that Reason about PeopleAnca Dragan [homepage]
The goal of my research is to enable robots to work with, around, and in support of people, autonomously producing behavior that reasons about both their function and their interaction with humans. I aim to develop a formal understanding of interaction that leads to algorithms which are informed by mathematical models of how humans interact with robots, enabling generalization across robot morphologies and interaction modalities. In this talk, I will focus on one specific instance of this agenda: autonomously generating motion for coordination during human-robot collaborative manipulation. Most motion in robotics is solely functional: industrial robots move to package parts, vacuuming robots move to suck dust, and personal robots move to clean up a dirty table. This type of motion is ideal when the robot is performing a task in isolation. Collaboration, however, does not happen in isolation, and demands that we move beyond solely functional motion. In collaboration, the robot's motion has an observer, watching and interpreting the motion – inferring the robot's intent from the motion, and anticipating the robot's motion based on its intent. My work integrates a mathematical model of these inferences into motion planning, so that the robot can generate motion that matches people's expectations and clearly conveys its intent. In doing so, I draw on action interpretation theory, Bayesian inference, constrained trajectory optimization, and interactive learning. The resulting motion not only leads to more efficient collaboration, but also increases the fluency of the interaction as defined through both objective and subjective measures. About the speaker:Anca Dragan is a new Assistant Professor at UC Berkeley's EECS Department. She completed her PhD in Robotics at Carnegie Mellon. She was born in Romania and received her B.Sc. in Computer Science from Jacobs University in Germany in 2009. Her research lies at the intersection of robotics, machine learning, and human-computer interaction: she works on algorithms that enable robots to seamlessly work with, around, and in support of people. Anca's research and her outreach activities with children have been recognized by the Intel Fellowship and by scholarships from Siebel, the Dan David Prize, and Google Anita Borg. |
Friday, October 9, 2015, 11:00AM
|
Theory and Practice: Crowdsourcing and Cost AllocationNick Mattei [homepage]
Modern technology enables computers and (by proxy) humans to communicate at distances and speeds previously unimaginable, connecting large numbers of agents over time and space. These groups of agents must make collective decisions, subject to constraints and preferences, in important settings including: item selection; resource or task allocation; and cost distribution. In CS, these topics fall into algorithmic game theory (AGT) computational social choice (ComSoc). Results in these areas have impact within (and use) AI, decision theory, optimization, recommender systems, data mining, and machine learning. Many of the key theoretical results in these areas are grounded on worst case assumptions about agent behavior or the availability of resources. Transitioning these theoretical results into practice requires data driven analysis and experiment. We detail two projects focused on applying theoretical results to real world decision making. First, we detail a novel mechanism for using crowd sourcing when selecting peers. Second, we show practical applications of cooperative game theory to the problem of dividing delivery costs to clients. Both of these projects leverage data to perform detailed experiments and resulted in deployable, open source decision tools. About the speaker:Nicholas Mattei is a senior researcher in the Optimization Research Group at NICTA and a conjoint lecturer at the University of New South Wales in Sydney, Australia. His research focuses on computational aspects of social choice, preference aggregation, and assignment; using computation to enable and augment human decision making. Along with Prof. Toby Walsh, he is the founder and maintainer of PrefLib: A Library for Preferences. He previously worked as a programmer and embedded electronics designer for nano-satellites at NASA Ames Research Center. He received his Ph.D from the University of Kentucky under the supervision of Prof. Judy Goldsmith in 2012. |
Friday, October 30, 2015, 11:00AM
|
Deep Learning for Decision Making and ControlSergey Levine [homepage]
A remarkable feature of human and animal intelligence is the ability to autonomously acquire new behaviors. My work is concerned with designing algorithms that aim to bring this ability to robots. A central challenge in this field is to learn behaviors with representations that are sufficiently general and expressive to handle the wide range of motion skills that are necessary for real-world applications, such as general-purpose household robots. These representations must also be able to operate on raw, high-dimensional inputs and outputs, such as camera images and joint torques. I will describe a class of guided policy search algorithms that tackle this challenge by transforming the task of learning control policies into a supervised learning problem, with supervision provided by simple, efficient trajectory-centric methods. I will show how this approach can be applied to a wide range of tasks, from locomotion and push recovery to robotic manipulation. I will also present new results on using deep convolutional neural networks to directly learn policies that combine visual perception and control, learning the entire mapping from rich visual stimuli to motor torques on a PR2 robot. I will conclude by discussing future directions in deep sensorimotor learning and how advances in this emerging field can be applied to a range of other domains.
About the speaker:Sergey Levine is a postdoctoral researcher working with Professor Pieter Abbeel at UC Berkeley, and a research scientist at Google. In spring 2016, he will be an Assistant Professor at the University of Washington. He completed his PhD in 2014 with Vladlen Koltun at Stanford University. His research focuses on robotics, machine learning, and computer graphics. In his PhD thesis, he developed a novel guided policy search algorithm for learning complex neural network control policies. In later work, this method enabled learning a range of robotic manipulation tasks, as well as end-to-end training of policies for perception and control. He has also developed algorithms for learning from demonstration, inverse reinforcement learning, and data-driven character animation. |
Friday, November 20, 2015, 11:00AM
|
Multiagent Simulation for Designing Social ServicesItsuki Noda [homepage]
Computer simulations of social phenomena will become the most
efficient tool to design and to improve social systems. Big data
and advancement of computational powers enable to handle large
scale social simulations in which a large number of human
activities are represented by behaviors of multiple intelligent
agent. We are conducting a project to establish multi-agent social
simulations and to apply them to actual real-world problems like
disaster mitigation, smart transportation systems, stable
economical systems, and so on. In this talk, I will show several
results of this project.
About the speaker:Itsuki Noda is a team leader of Service Design Assist Research Team of Center for Service Research, National Institute of Advanced Industrial Science and Technology (AIST), Japan. He received the B.E., M.E. and Ph.D., degrees in electrical engineering from Kyoto University, Kyoto, Japan, in 1987, 1989, and 1995, respectively. He was a visiting researcher of Stanford University in 1999, and worked as a staff of Council of Science and Technology Policy of Japanese government in 2003. He was a founding member of RoboCup and promoted Simulation League since 1995. RoboCup is a research competition and symposium on robotics and artificial intelligence, and is held the international competitions every year. The Simulation League becomes a standard problem on researchs multi-agent simulation domain and used world-wide. Now, he is the president of RoboCup Federation. |
Monday, November 23, 2015, 2:00PM
|
Learning Deep Object Detectors from Synthetic 3D CAD ModelsKate Saenko [homepage]
Supervised deep convolutional neural networks (DCNN) have significantly advanced object detection in recent years. However, the very large amount of bounding-box annotated training data they require is difficult to obtain for many objects. On the other hand, crowdsourced 3D CAD models are becoming easily accessible online, and can potentially generate an infinite number of training images for almost any object category! We show that augmenting the training data of contemporary Deep Convolutional Neural Net (DCNN) models with such synthetic data can be effective, especially when real training data is limited or not well matched to the target domain. Most freely available CAD models capture 3D shape but are often missing other low level cues, such as realistic object texture, pose, or background. In a detailed analysis, we use synthetic CAD-rendered images to probe the ability of DCNN to learn without these cues, with surprising findings. In particular, we show that when the DCNN is fine-tuned on the target detection task, it exhibits a large degree of invariance to missing low-level cues. We analyse conditions under which fine-tuning on synthetic data improves performance on real images. About the speaker:Kate Saenko is an Assistant Professor of Computer Science at the University of Massachusetts Lowell. She received her PhD from MIT, followed by postdoctoral work at UC Berkeley and Harvard. Her research spans the areas of computer vision, machine learning, speech recognition, and human-robot interfaces. Dr Saenko’s current research interests include domain adaptation for object recognition and joint modeling of language and vision. |
Friday, December 4, 2015, 11:00AM
|
Scaling Up Object DetectionOlga Russakovsky [homepage]
Hundreds of billions of photographs are uploaded on the web each year. An important step towards automatically analyzing the content of these photographs is building computer vision models that can recognize and localize all the depicted objects. Traditionally, work on scaling up object recognition has focused on algorithmic improvements, e.g., building more efficient or more powerful models. However, I will argue that data plays at least as big a role: effectively collecting and annotating the right data is a critical component of scaling up object detection. The first part of the talk will be about constructing an object detection dataset (as part of the ImageNet Large Scale Visual Recognition Challenge) that is an order of magnitude larger than previous datasets such as the PASCAL VOC. I will discuss some of the decisions we made in designing this benchmark as well as some of our crowd engineering innovations. The availability of this large-scale data gave us as a field an unprecedented opportunity to work on designing algorithms for scalable and diverse object detection. It also allows for thorough analysis to understand the current algorithmic shortcomings and to focus the next round of algorithmic improvements. In the second part of the talk, I will bring together the insights from large-scale data collection and from recent algorithmic innovations into a principled human-in-the-loop framework for image understanding. This approach can be used both for reducing the cost of large-scale detailed dataset annotation efforts as well as for effectively understanding a single target image.
About the speaker:Olga Russakovsky (http://cs.cmu.edu/~orussako) is a postdoctoral research fellow at Carnegie Mellon University. She recently completed a PhD in computer science at Stanford University advised by Prof. Fei-Fei Li. Her research interests are in computer vision and machine learning, specifically focusing on large-scale object detection and recognition. She was the lead organizer of the ImageNet Large Scale Visual Recognition Challenge (http://image-net.org/challenges/LSVRC) for two years, which was featured in the New York Times and MIT Technology Review. She organized multiple workshops and tutorials at premier computer vision conferences, including helping pioneer the “Women in Computer Vision” workshop at CVPR’15. She founded and directs the Stanford AI Laboratory’s outreach camp SAILORS (http://sailors.stanford.edu, featured in Wired) designed to expose high school students in underrepresented populations to the field of AI. |
Friday, January 22, 2016, 11:00AM
|
Situated Learning and Understanding of Natural LanguageYoav Artzi [homepage]
Robust language understanding systems have the potential to transform how we interact with computers. However, significant challenges in automated reasoning and learning remain to be solved before we achieve this goal. To accurately interpret user utterances, for example when instructing a robot, a system must jointly reason about word meaning, grammatical structure, conversation history and world state. Additionally, to learn without prohibitive data annotation costs, systems must automatically make use of weak interaction cues for autonomous language learning. In this talk, I will present a framework that uses situated interactions to learn to map sentences to rich, logical meaning representations. The approach jointly induces the structure of a complex natural language grammar and estimates its parameters, while relying on various learning cues, such as easily gathered demonstrations and even raw conversations without any additional annotation effort. It achieves state-of-the-art performance on a number of tasks, including robotic interpretation of navigational directions and learning to understand user utterances in dialog systems. Such an approach, when integrated into complete systems, has the potential to achieve continuous, autonomous learning by participating in actual interactions with users. About the speaker:Yoav Artzi is an Assistant Professor in the Department of Computer Science and Cornell Tech at Cornell University. His research interests are in the intersection of natural language processing and machine learning. In particular, he focuses on designing latent variable learning algorithms that recover rich representations of linguistic meaning for situated natural language understanding. He received the best paper award in EMNLP 2015, his B.Sc. summa cum laude from Tel Aviv University, and his Ph.D. from the University of Washington. |
Monday, April 4, 2016, 11:00AM
|
Using Images to Predict ConsequencesGarrett Warnell [homepage]
In this talk, I will discuss recent work pertaining to two different ongoing projects I am involved with at the US Army Research Laboratory. In the first part, motivated by the goal of autonomous visual exploration, I will consider the problem of extending computational visual saliency techniques to the robotics domain. Specifically, I propose a new method we have developed for computing visual saliency for sets of images collected using a pan-tilt-zoom camera. In the second part of the talk, motivated by increasing the operational tempo of a ground robot in new environments, I consider the problem of predicting physical-state-estimation error. Specifically, I will describe a novel online learning algorithm that we have proposed using a statistical model of error that is able to provide enough expressive power to enable prediction directly from motion control signals and low-level visual features.
About the speaker:Garrett Warnell is a Research Scientist in the Computational and Information Sciences Directorate at the US Army Research Laboratory in Adelphi, Maryland. His research interests are broadly in the areas of machine learning, computer vision, and robotics. He received a BS in computer engineering and a BS in mathematics from Michigan State University in 2009, and a MS and PhD in electrical engineering from the University of Maryland, College Park in 2013 and 2014, respectively. |
Friday, April 15, 2016, 11:00AM
|
Off-Policy Policy EvaluationPhilip Thomas [homepage]
In this talk I will present some exciting recent advances in "off-policy policy evaluation", which is the problem of predicting how well a new policy will work given historical data that was generated by a different policy. The ability to evaluate a new policy from historical data is important for applications where the deployment of a bad policy can be dangerous or costly. After presenting the Doubly Robust estimator recently proposed by Nan Jiang and Lihong Li, I will explore methods for further improving its accuracy, culminating with the "MAGIC" estimator, which I recently proposed in collaboration with Emma Brunskill. Intuitively, the MAGIC estimator leverages a new way for combining model-based and model-free estimates, which is based on the idea of complex returns, like the lambda-return. Empirical results suggest that the MAGIC estimator usually performs at least as well as the best previously existing algorithm, and often orders of magnitude better (in terms of mean squared error).
About the speaker:Philip Thomas is interested in artificial intelligence research, primarily reinforcement learning research. His past work has focused on predicting the performance of new solutions (policies) from historical data, natural gradient methods and their application to policy search, and trying to improve the biological plausibility of policy gradient algorithms by estimating gradients without the need for backpropagation.Philip received his B.S. and M.S. in computer science from Case Western Reserve University in 2008 and 2009 respectively, where Michael Branicky was his adviser. He then completed his PhD at UMass Amherst under the supervision of Andrew Barto. He is now a postdoctoral researcher at CMU, where he is supervised by Emma Brunskill. |
Monday, April 25, 2016, 11:00AM
|
Deep learning with disentangled representationsHonglak Lee [homepage]
Over the recent years, deep learning has emerged as a powerful method for learning feature representations from complex input data, and it has been greatly successful in computer vision, speech recognition, and language modeling. The recent successes typically rely on a large amount of supervision (e.g., class labels). While many deep learning algorithms focus on a discriminative task and extract only task-relevant features that are invariant to other factors, complex sensory data is often generated from intricate interaction between underlying factors of variations (for example, pose, morphology and viewpoints for 3d object images). In this work, we tackle the problem of learning deep representations that disentangle underlying factors of variation and allow for complex reasoning and inference that involve multiple factors. Specifically, we develop deep generative models with higher-order interactions among groups of hidden units, where each group learns to encode a distinct factor of variation. We present several successful instances of deep architectures and their learning methods, including supervised and weakly-supervised setting. Our models achieve strong performance in emotion recognition, face verification, data-driven modeling of 3d objects, and video game prediction. I will also present other related ongoing work.
About the speaker:Honglak Lee is an Assistant Professor of Computer Science and Engineering at the University of Michigan, Ann Arbor. He received his Ph.D. from Computer Science Department at Stanford University in 2010, advised by Prof. Andrew Ng. His research focuses on deep learning and representation learning, which spans over unsupervised and semi-supervised learning, supervised learning, transfer learning, structured prediction, graphical models, and optimization. His methods have been successfully applied to computer vision and other perception problems. He received best paper awards at ICML and CEAS. He has served as a guest editor of IEEE TPAMI Special Issue on Learning Deep Architectures, as well as area chairs of ICML, NIPS, ICCV, AAAI, IJCAI, and ICLR. He received the Google Faculty Research Award (2011), NSF CAREER Award (2015), and was selected by IEEE Intelligent Systems as one of AI's 10 to Watch (2013) and a fellow by Alfred P. Sloan Foundation (2016). |