-------------------------------------Past Events ---------------------------------------------
IEEE-Dayton Section AES Society Meeting:
TITLE: Sensing and Decision Making amongst Networked Social Sensors
SPEAKER: Professor Vikram Krishnamurthy, Cornell University
DATE: Thursday, June 28, 2018
TIME: 11:30 AM - 12:30 PM
PLACE: Neuroscience Engineering Collaboration Building, Room 101, Wright State University
RSVP REQUESTED: Justin Metcalf - https://events.vtools.ieee.org/m/173982
Meeting/Presentation is open to all
Abstract: This talk discusses how humans interact over a social network and make decisions based on sensor information. Humans can be viewed as social sensors that input information to a social network. The interaction of social sensors present several challenges from a statistical signal processing viewpoint: sensors interact with and influence other social sensors resulting in herding behaviour. Second, due to privacy concerns, social sensors reveal quantized decisions (ratings, recommendations). Third, social sensors are risk averse decision makers with anticipatory emotions. This talk describes mathematical models for how social sensors interact over a social network, how social sensor decision-making can result in herding behaviour, and how herding can be mitigated by providing incentives to individual sensors. We will also discuss novel methods to poll social networks based on expectation polling and the friendship paradox. The seminar draws from ideas in statistical signal processing and behavioral economics.
Bio: Dr. Krishnamurthy joined Cornell in 2016. Previously, he was a Canada Research Chair professor at University of British Columbia. Dr. Krishnamurthy is a Fellow of the IEEE. He was awarded an honorary doctorate from Royal Institute of Technology, Sweden. He was a Distinguished Lecturer for IEEE Signal Processing Society, and Editor in Chief of IEEE Journal Selected Topics in Signal Processing. He is also with the Center for Applied Math at Cornell.
2017 ATR Center Seminars
Thursday June 8, 3pm
|Clustering Consistently: Densities and Graphons
|Thursday June 15, 3pm
|Fast and Robust Phase Retrieval with Provable Guarantees
|Thursday June 29, 3pm
|Compressive sensing tutorial
|Thursday July 6, 3pm
|Message passing algorithms
|Tuesday July 18, 3pm
|Computational imaging -- The Co-design of optics and algorithms for tractable inference
All of above seminars are scheduled to be held at Wright State University in Medical Sciences Bldg, Room 120.
Clustering Consistently: Densities and Graphons Justin Eldridge
In what sense does a clustering algorithm recover the "correct" clusters? In a statistical setting where the data are drawn from a probability distribution, we might naturally expect a "correct" algorithm to recover the cluster structure of the underlying distribution. In this talk, I will discuss two important and popular cases: when we wish to cluster 1) points sampled from a density, and 2) the nodes of a graph generated from a graphon -- a very powerful random graph model of much recent interest. I will show that in both cases the clusters of the distribution form a tree and argue that a natural goal of clustering is to produce an empirical tree which converges to the ideal in the limit of large data. To this end, I will present a merge distortion metric for measuring the distance between the output of a clustering algorithm and the ideal tree of the distribution. I will prove that, in both cases, efficient algorithms exist which converge to the ideal tree in the sense of the merge distortion metric, such that these clustering methods are "correct" in a precise sense.
Fast and Robust Phase Retrieval with Provable Guarantees Yuejie Chi
The goal of phase retrieval is to recover a signal of interest from only the amplitudes of its linear measurements, e.g. from the modulus of a diffracted wave. It arises in a variety of important applications such as X-ray crystallography, coherent diffraction imaging and SAR imaging, where it is difficult to measure the phase information. This talk will discuss the classical Gerchberg-Saxton algorithm as well as recent algorithmic approaches based on gradient descent using an amplitude-based loss function, and describe their theoretical performance guarantees under a certain random measurement model. In particular, it is demonstrated that they converge at a linear rate to the ground truth when initialized properly at a near-optimal sample complexity. We will also describe stochastic and robust variants of the gradient descent approach when the measurements are either accessed in a mini-batch manner or possibly corrupted by arbitrary outliers. Finally, numerical simulations will be provided to demonstrate the efficacy of the discussed approaches.
Computational Imaging --- The Co-design of optics and algorithms for tractable inference Aswin Sankaranarayanan
The basic design of the camera — a lens and an imaging sensor — has remain unchanged for many centuries. While this design has been perfected for photography, there are many current and upcoming application domains where it is woefully inadequate. I will discuss a few examples that demonstrate how the co-design of optics and processing, often referred to as computational imaging, can supplant the limitations of conventional imaging designs. I will discuss examples of novel imaging designs and associated processing techniques that provide capabilities that far exceed those of the conventional design while allowing for tractable inference to problems that are otherwise hard to solve.