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 ]
Friday, November 3, 2017, 11:00AM
|
Deep Learning, Where are you going?Kyunghyun Cho [homepage]
There are three axes along which advances in machine learning and deep learning
happen. They are (1) network architectures, (2) learning algorithms and (3)
spatio-temporal abstraction. In this talk, I will describe a set of research
topics I've pursued in each of these axes. For network architectures, I will
describe how recurrent neural networks, which were largely forgotten during 90s
and early 2000s, have evolved over time and have finally become a de facto
standard in machine translation. I continue on to discussing various learning
paradigms, how they related to each other, and how they are combined in order
to build a strong learning system. Along this line, I briefly discuss my latest
research on designing a query-efficient imitation learning algorithm for
autonomous driving. Lastly, I present my view on what it means to be a
higher-level learning system. Under this view each and every end-to-end
trainable neural network serves as a module, regardless of how they were
trained, and interacts with each other in order to solve a higher-level task.
I will describe my latest research on trainable decoding algorithm as a first
step toward building such a framework.
About the speaker:Kyunghyun Cho is an assistant professor of computer science and data science at New York University. He was a postdoctoral fellow at University of Montreal until summer 2015, and received PhD and MSc degrees from Aalto University early 2014. He tries best to find a balance among machine learning, natural language processing and life, but often fails to do so. |
Friday, December 1, 2017, 11:00AM
|
Bootstrapping Cognitive SystemsKenneth D. Forbus [homepage]
Annotation, even by crowdsourcing, is a trap: People learn without others having privileged access to their mental states. Today’s ML systems require a cadre of technical experts to train them, using massively more data than people require, whereas people manage their own learning processes over a lifetime. Understanding how to build cognitive systems that can learn well from small amounts of data, expressed in forms natural to people, and able to manage their own learning over extended periods would be a revolutionary advance. In the Companion cognitive architecture, we are exploring ways to do this, using a combination of analogy and relational representations. This talk will describe several recent advances we have made, including learning human behavior from Kinect data, analogical chaining for commonsense reasoning, and co-learning of language disambiguation and reasoning using unannotated data. Ideas for scaling an analogical approach to cognitive systems to human-sized knowledge bases and potential applications along the way will also be discussed.
About the speaker:Kenneth D. Forbus is the Walter P. Murphy Professor of Computer Science and Professor of Education at Northwestern University. He received his degrees from MIT (Ph.D. in 1984). His research interests include qualitative reasoning, analogical reasoning and learning, spatial reasoning, sketch understanding, natural language understanding, cognitive architecture, reasoning system design, intelligent educational software, and the use of AI in interactive entertainment. He is a Fellow of the Association for the Advancement of Artificial Intelligence, the Cognitive Science Society, and the Association for Computing Machinery. He has received the Humboldt Award and has served as Chair of the Cognitive Science Society. |
Wednesday, December 13, 2017, 11:00AM
|
Teaching Social Skills with Social RobotsBrian Scassellati [homepage]
In the last decade, there has been a slowly growing interaction between robotics researchers and clinicians to look at the viability of using robots as a tool for enhancing therapeutic and diagnostic options for individuals with autism spectrum disorder. While much of the early work in using robots for autism therapy lacked clinical rigor, new research is beginning to demonstrate that robots improve engagement and elicit novel social behaviors from people (particularly children and teenagers) with autism. However, why robots in particular show this capability, when similar interactions with other technology or with adults or peers fails to show this response, remains unknown. This talk will present some of the most recent evidence showing robots eliciting social behavior from individuals with autism and discuss some of the mechanisms by which these effects may be generated.
About the speaker:Brian Scassellati is a Professor of Computer Science, Cognitive Science, and Mechanical Engineering at Yale University and Director of the NSF Expedition on Socially Assistive Robotics. His research focuses on building embodied computational models of human social behavior, especially the developmental progression of early social skills. Using computational modeling and socially interactive robots, his research evaluates models of how infants acquire social skills and assists in the diagnosis and quantification of disorders of social development (such as autism). His other interests include humanoid robots, human-robot interaction, artificial intelligence, machine perception, and social learning. Dr. Scassellati received his Ph.D. in Computer Science from the Massachusetts Institute of Technology in 2001. His dissertation work (Foundations for a Theory of Mind for a Humanoid Robot) with Rodney Brooks used models drawn from developmental psychology to build a primitive system for allowing robots to understand people. His work at MIT focused mainly on two well-known humanoid robots named Cog and Kismet. He also holds a Master of Engineering in Computer Science and Electrical Engineering (1995), and Bachelors degrees in Computer Science and Electrical Engineering (1995) and Brain and Cognitive Science (1995), all from MIT. Dr. Scassellati's research in social robotics and assistive robotics has been recognized within the robotics community, the cognitive science community, and the broader scientific community. He was named an Alfred P. Sloan Fellow in 2007 and received an NSF CAREER award in 2003. His work has been awarded five best-paper awards. He was the chairman of the IEEE Autonomous Mental Development Technical Committee from 2006 to 2007, the program chair of the IEEE International Conference on Development and Learning (ICDL) in both 2007 and 2008, and the program chair for the IEEE/ACM International Conference on Human-Robot Interaction (HRI) in 2009. Descriptions of his recent work have been published in the Wall Street Journal (reprinted here), the New York Times Sunday Magazine, Popular Science, New Scientist, the APA Monitor on Psychology, SEED Magazine, and NPR's All Things Considered. |
Friday, January 12, 2018, 11:00AM
|
Marine Robotics: Planning, Decision Making, and LearningGeoffrey A. Hollinger [homepage]
Underwater gliders, propeller-driven submersibles, and other marine robots are increasingly being tasked with gathering information (e.g., in environmental monitoring, offshore inspection, and coastal surveillance scenarios). However, in most of these scenarios, human operators must carefully plan the mission to ensure completion of the task. Strict human oversight not only makes such deployments expensive and time consuming but also makes some tasks impossible due to the requirement for heavy cognitive loads or reliable communication between the operator and the vehicle. We can mitigate these limitations by making the robotic information gatherers semi-autonomous, where the human provides high-level input to the system and the vehicle fills in the details on how to execute the plan. These capabilities increase the tolerance for operator neglect, reduce deployment cost, and open up new domains for information gathering.
In this talk, I will show how a general framework that unifies information theoretic optimization and physical motion planning makes semi-autonomous information gathering feasible in marine environments. I will leverage techniques from stochastic motion planning, adaptive decision making, and deep learning to provide scalable solutions in a diverse set of applications such as underwater inspection, ocean search, and ecological monitoring. The techniques discussed here make it possible for autonomous marine robots to “go where no one has gone before,” allowing for information gathering in environments previously outside the reach of human divers.
About the speaker:Geoffrey A. Hollinger is an Assistant Professor in the School of Mechanical, Industrial & Manufacturing Engineering at Oregon State University. His current research interests are in adaptive information gathering, distributed coordination, and learning for autonomous robotic systems. He has previously held research positions at the University of Southern California, Intel Research Pittsburgh, University of Pennsylvania’s GRASP Laboratory, and NASA's Marshall Space Flight Center. He received his Ph.D. (2010) and M.S. (2007) in Robotics from Carnegie Mellon University and his B.S. in General Engineering along with his B.A. in Philosophy from Swarthmore College (2005). He is a recent recipient of the 2017 Office of Naval Research Young Investigator Program (YIP) award. |
Friday, January 26, 2018, 11:00AM
|
A Distributional Perspective on Reinforcement LearningMarc Bellemare [homepage]
This talk will present an overview of our recent research on distributional reinforcement learning. Our starting point is our recent ICML paper, in which we argued for the fundamental importance of the value distribution: the distribution of random returns received by a reinforcement learning agent. This is in contrast to the common approach, which models the expectation of this return, or value. Back then, we were able to design a new algorithm that learns the value distribution through a TD-like bootstrap process and achieved state-of-the-art performance on games from the Arcade Learning Environment (ALE). However, this left open the question as to why the distributional approach should perform better at all. We've since delved deeper into what makes distributional RL work: first by improving the original using quantile regression, which directly minimizes the Wasserstein metric; and second by unearthing surprising connections between the original C51 algorithm and the distant cousin of the Wasserstein metric, the Cramer distance.
About the speaker:As humans, we spend our daily lives interacting with the unknown, from ordering off a restaurant menu to learning the ropes at a new job. From an artificial intelligence perspective, we are generally competent agents. I am interested in understanding which algorithms may support this general competency in artificial agents and developing the right environments, simulated or otherwise, to foster and study general competency. I am a research scientist at Google Brain in Montréal, Canada focusing on the reinforcement learning effort there. From 2013 to 2017 I was at DeepMind in the UK. I received my Ph.D. from the University of Alberta working with Michael Bowling and Joel Veness. My research lies at the intersection of reinforcement learning and probabilistic prediction. I'm also interested in deep learning, generative modelling, online learning, and information theory. |
Friday, February 9, 2018, 11:00AM
|
Text Generation in an End-to-End WorldSasha Rush [homepage]
Progress in neural machine translation has led to optimism for text generation tasks such as summarization and dialogue, but compared to translation it has been more difficult to quantify the successes and remaining challenges. In this talk, I will survey some of the recent advances in neural generation, and present a successful implementation of these techniques in the recent E2E NLG data-to-text challenge (Gehrmann et al, 2018). Despite this early progress, though, further experiments show that end-to-end models fail to scale to more realistic data-to-document scenarios, and need further improvements to robustly capture higher-level structure and content selection (Wiseman et al, 2017). I will end by discussing recent work in the stranger area of unsupervised NLG where we show interesting results in neural text style transfer using a GAN-based autoencoder (Zhao et al 2017).
About the speaker:Alexander "Sasha" Rush is an assistant professor at Harvard University. His research interest is in ML methods for NLP with a focus on deep learning methods for applications in machine translation, data and document summarization, and diagram processing, as well as the development of the OpenNMT translation system. His past work focused on structured prediction and combinatorial optimization for NLP. Sasha received his PhD from MIT supervised by Michael Collins and was a postdoc at Facebook NY under Yann LeCun. His work has received four research awards at major NLP conferences. |
Friday, March 23, 2018, 11:00AM
|
Natural Language Learning for Human-Robot CollaborationMatt Walter [homepage]
Natural language promises an efficient and flexible means for humans to communicate with robots, whether they are assisting the physically or cognitively impaired, or performing disaster mitigation tasks as our surrogates. Recent advancements have given rise to robots that are able to interpret natural language commands that direct object manipulation and spatial navigation. However, most methods require prior knowledge of the metric and semantic properties of the objects and places that comprise the robot's environment. In this talk, I will present our work that enables robots to successfully follow natural language navigation instructions within novel, unknown environments. I will first describe a method that treats language as a sensor, exploiting information implicit and explicit in the user's command to learn distributions over the latent spatial and semantic properties of the environment and over the robot's intended behavior. The method then learns a belief space policy that reasons over these distributions to identify suitable navigation actions. In the second part of the talk, I will present an alternative formulation that represents language understanding as a multi-view sequence-to-sequence learning problem. I will introduce an alignment-based neural encoder-decoder architecture that translates free-form instructions to action sequences based on images of the observable world. Unlike previous methods, this architecture uses no specialized linguistic resources and can be trained in a weakly supervised, end-to-end fashion, which allows for generalization to new domains. Time permitting, I will then describe how we can effectively invert this model to enable robots to generate natural language utterances. I will evaluate the efficacy of these methods on a combination of benchmark navigation datasets and through demonstrations on a voice-commandable wheelchair.
About the speaker:Matthew Walter is an assistant professor at the Toyota Technological Institute at Chicago. His interests revolve around the realization of intelligent, perceptually aware robots that are able to act robustly and effectively in unstructured environments, particularly with and alongside people. His research focuses on machine learning-based solutions that allow robots to learn to understand and interact with the people, places, and objects in their surroundings. Matthew has investigated these areas in the context of various robotic platforms, including autonomous underwater vehicles, self-driving cars, voice-commandable wheelchairs, mobile manipulators, and autonomous cars for (rubber) ducks. Matthew obtained his Ph.D. from the Massachusetts Institute of Technology, where his thesis focused on improving the efficiency of inference for simultaneous localization and mapping. |
Friday, March 30, 2018, 11:00AM
|
New Frontiers in Imitation LearningYisong Yue [homepage]
In this talk, I will describe recent and ongoing work in developing principled imitation learning approaches that can exploit structural interdependencies in the state/action space. Compared to conventional imitation learning, these approaches can achieve orders-of-magnitude improvements in learning rate or accuracy, or both. These approaches are showcased on a wide range of (often commercially deployed) applications, including modeling professional sports, laboratory animals, speech animation, and expensive computational oracles.
About the speaker:Yisong Yue is an assistant professor in the Computing and Mathematical Sciences Department at the California Institute of Technology. He was previously a research scientist at Disney Research. Before that, he was a postdoctoral researcher at Carnegie Mellon University. He received a Ph.D. from Cornell University and a B.S. from the University of Illinois at Urbana-Champaign. Yisong's research interests lie primarily in the theory and application of statistical machine learning. His research is largely centered around developing learning approaches that can characterize structured and adaptive decision-making settings. In the past, his research has been applied to information retrieval, recommender systems, text classification, learning from rich user interfaces, analyzing implicit human feedback, data-driven animation, behavior analysis, sports analytics, policy learning in robotics, and adaptive planning & allocation problems. |
Friday, April 6, 2018, 11:00AM
|
Event Coreference Resolution by Iteratively Unfolding Inter-dependencies among EventsRuihong Huang [homepage]
Recognizing all references to the same event within a document as well as across documents is vital for information aggregation, storyline generation and many NLP applications, such as event detection and tracking, question answering and text summarization. While it can be more risky and require additional evidence to link event mentions from two distinct documents, resolving event coreference links within a document is equally challenging due to dissimilar event word forms, incomplete event arguments (e.g., event participants, time and location) and distinct contexts. In this talk, I will present our recent work on event coreference resolution that tackles the conundrum and gradually builds event clusters by exploiting inter-dependencies among event mentions within the same chain as well as across event chains in an iterative joint inference approach. I will then briefly discuss various applications of event coreference resolution and the future directions.
About the speaker:Ruihong Huang is an Assistant Professor in the Computer Science and Engineering Department at Texas A&M University, College Station. Dr. Huang received her PhD in computer science at the University of Utah. She joined TAMU in Fall 2015 after she completed a Postdoc at Stanford University. Her research is mainly on computational linguistics and machine learning, with special research interests on information extraction, discourse and semantics. Her research spans from extracting propositional facts from texts to studying extra-propositional aspects of meanings and various subjectivities. |
Friday, April 27, 2018, 11:00AM
|
Domain Randomization for Transferring Neural Networks from Simulation to the Real WorldJosh Tobin [homepage]
Robotic data can be expensive to collect and hard to label. So to apply data-intensive machine learning techniques like deep learning to problems in robotics, we would like to be able to perform most of our learning on cheap and easy to label data from physics simulators. However, models learned in simulation often perform badly on physical robots due to the 'reality gap' that separates real-world robotics from even the best-tuned simulators. In this talk we will discuss a simple and surprisingly powerful technique for bridging the reality gap called domain randomization. Domain randomization involves massively randomizing non-essential aspects of the simulator so that the model is forced to learn to ignore them. We will talk about applications of this idea in visual domains for object pose estimation and for generating synthetic objects for robotic grasping.
About the speaker:Josh Tobin is a Research Scientist at OpenAI and a PhD student in Computer Science at UC Berkeley working under the supervision of Professor Pieter Abbeel. Josh's research focuses on applying deep learning to problems in robotic perception and control, with a particular concentration on deep reinforcement learning, transfer learning, domain adaptation, and generative models. |
Monday, May 7, 2018, 11:00AM
|
Robot Learning through Motion and InteractionDave Held [homepage]
Robots today are typically confined to operate in relatively simple, controlled environments. One reason for these limitation is that current methods for robotic perception and control tend to break down when faced with occlusions, viewpoint changes, poor lighting, unmodeled dynamics, and other challenging but common situations that occur when robots are placed in the real world. I argue that, in order to handle these variations, robots need to learn to understand how the world changes over time: how the environment can change as a result of the robot’s own actions or from the actions of other agents in the environment. I will show how we can apply this idea of understanding changes to a number of robotics problems, such as object segmentation, tracking, and velocity estimation for autonomous driving as well as perception and control for various object manipulation tasks. By learning how the environment can change over time, we can enable robots to operate in the complex, cluttered environments of our daily lives.
About the speaker:David Held is an assistant professor in the Robotics Institute at CMU, working on robotic perception for autonomous driving and object manipulation. Prior to coming to CMU, he was a post-doctoral researcher at U.C. Berkeley where he worked on deep reinforcement learning, and he completed his Ph.D. in Computer Science at Stanford University where he developed methods for perception for autonomous vehicles. David has also worked as an intern on Google’s self-driving car team. David has a B.S. and M.S. in Mechanical Engineering at MIT. David is a recipient of the Google Faculty Research Award in 2017. |