IEEE CIS Page IEEE SCV Page IEEE-Egrid IEEE-Egrid

When your big data seems too small: accurate inferences beyond the empirical distribution

Gregory Valiant, Stanford University

Event Co-Sponsors:

Date & Time: Thursday, August 23, 2018, 6:00 PM – 8:00 PM PDT

Location: Intel SC12, 3600 Juliette Ln, Santa Clara, CA 95054

Registration Link: (Mandatory) : EventBrite Link

Registration Fee: IEEE CIS members: free
         Students - $3 (Register at Door $3)
         IEEE (non-CIS) members - $7 donation (Register at Door $10)
         Non-members - $10 (Register at Door $15)

Share Event with your Colleagues:

LinkedIn Email Facebook Twitter
prof_GregoryValiant

Abstract: We discuss several problems related to the general challenge of making accurate inferences about a complex phenomenon, in the regime in which the amount of available data (i.e the sample size) is too small for the empirical distribution of the samples to be an accurate representation of the phenomenon in question. We show that for several fundamental and practically relevant settings, including estimating the covariance structure of a high-dimensional distribution, and learning a population of distributions given few data points from each individual, it is possible to ``denoise'' the empirical distribution significantly. We will also discuss the problem of estimating the ``learnability'' of a dataset: given too little labeled data to train an accurate model, we show that it is often possible to estimate the extent to which a good model exists. Framed differently, even in the regime in which there is insufficient data to learn, it is possible to estimate the performance that could be achieved if additional data (drawn from the same data source) were obtained. Our results, while theoretical, have a number of practical applications, and we also discuss some of these applications.

Biography: Gregory Valiant is an assistant professor of Computer Science at Stanford University. His current research interests span algorithms, statistics, and machine learning, with an emphasis on developing algorithms and information theoretic lower bounds for a variety of fundamental data-centric tasks. Recently, this work has also included questions of how to robustly extract meaningful information from untrusted datasets that might contain a significant fraction of corrupted or arbitrarily biased data points. Prior to joining Stanford, Gregory completed his PhD at UC Berkeley in 2012, and was a postdoctoral researcher at Microsoft Research, New England. He has received several honors, including the ACM Dissertation Award Honorable Mention, NSF Career Award, and Sloan Foundation Fellowship.

On the Role of Structure in Learning for Robot Manipulation

Jeannette Bohg, Stanford University

Event Co-Sponsors:

Date & Time: Thursday, September 20, 2018, 6:00 PM – 8:00 PM PDT

Location: Intel SC12, 3600 Juliette Ln, Santa Clara, CA 95054

Directions: Intel-SC12-Auditorium.pdf

Registration Link: (Mandatory) : EventBrite Link

Registration Fee: IEEE CIS members: free
         Students - $3 (Register at Door $3)
         IEEE (non-CIS) members - $7 donation (Register at Door $10)
         Non-members - $10 (Register at Door $15)

Share Event with your Colleagues:

LinkedIn Email Facebook Twitter
prof_JeannetteBohg

Abstract: Recent approaches in robotics follow the insight that perception is facilitated by interaction with the environment. First, this creates a rich sensory signal that would otherwise not be present. Second, knowledge of the sensory dynamics upon interaction allows prediction and decision-making over a longer time horizon. To exploit these benefits of Interactive Perception for capable robotic manipulation, a robot requires both: methods for processing rich, sensory feedback and feedforward predictors of the effect of physical interaction. In the first part of this talk, I will present a method for motion-based segmentation of an unknown number of simultaneously moving objects. The underlying model estimates dense, per-pixel scene flow that is then followed by clustering in motion trajectory space. We show how this outperforms state-of-the-art in scene flow estimation and multi-object segmentation. In the second part, I will present a method for predicting the effect of physical interaction with objects in the environment. The underlying model combines an analytical physics model and a learned perception part. In extensive experiments, we show how this hybrid model outperforms purely learned models in terms of generalisation. In both projects, we found that introducing structure greatly reduces training data, eases learning and provides extrapolation. Based on these findings, I will discuss the role of structure in learning for robot manipulation.

Biography: Jeannette Bohg is an Assistant Professor of Computer Science at Stanford University. She was a group leader at MPI until September 2017 and remains affiliated as a guest researcher. Her research focuses on perception for autonomous robotic manipulation and grasping. She is specifically interesting in developing methods that are goal-directed, real-time and multi-modal such that they can provide meaningful feedback for execution and learning. Before joining the Autonomous Motion lab in January 2012, Jeannette Bohg was a PhD student at the Computer Vision and Active Perception lab (CVAP) at KTH in Stockholm. Her thesis on Multi-modal scene understanding for Robotic Grasping was performed under the supervision of Prof. Danica Kragic. She studied at Chalmers in Gothenburg and at the Technical University in Dresden where she received her Master in Art and Technology and her Diploma in Computer Science, respectively.