SUNMONTUEWEDTHUFRISAT
Events for March 03, 2006
-
Delay, feedback, and the price of ignorance
Fri, Mar 03, 2006 @ 11:00 AM - 12:00 PM
Ming Hsieh Department of Electrical and Computer Engineering
Conferences, Lectures, & Seminars
SPEAKER: Dr. Anant Sahai, UC BerkeleyABSTRACT: In 1959, Shannon made a profound comment:"[The duality between source and channel coding] can be pursued further and is related to a duality between past and future and the notions of control and knowledge. Thus we may have knowledge of the past and cannot control it; we may control the future but have no knowledge of it."This comment cannot be understood in the traditional block-code setting and as a result, has remained entirely mysterious. To understand it, we must step back and consider end-to-end delay, since delay is what fundamentally allows the exploitation of the laws of large numbers to give reliability.In channel coding, we show that while feedback often does not improve fixed block-length reliability functions, it can significantly improve the reliability with respect to fixed delay! (Contrary to a "theorem" by Pinsker claiming otherwise.) A new bound, that we call the "focusing bound," allows us to calculate the limit of what is possible when the encoder is not ignorant of the channel's past behavior. In source coding, the price of ignorance is demonstrated by considering what happens when receiver side-information is withheld from the transmitter. Block-codes perform equally poorly, but nonblock codes can use side-information to dramatically improve the fixed-delay error exponent. Furthermore, a closer look at the dominant error events for these cases gives Shannon's otherwise cryptic comment a precise interpretation.These results suggest that the traditional information theoretic recommendation of using messages as big as possible is flawed as far as architectural guidance is concerned. When encoders are not ignorant, messages should be as *small* as possible while avoid integer effects, and queueing ideas should be employed to do appropriate flow control, even when facing hard end-to-end latency constraints.BIO: Anant Sahai received the B.S. degree in EECS in 1994 from U.C. Berkeley, and both his M.S. and Ph.D. degrees in EECS from the Massachusetts Institute of Technology, in 1996 and 2001, respectively. In 2001, he developed adaptive signal processing algorithms for software radio GPS at the startup Enuvis in South San Francisco. He joined the EECS department at Berkeley as an Assistant Professor in 2002. His current research interests are in information theory and wireless communication, particularly the area of opportunistic spectrum reuse by cognitive radios.Host: Professor Urbashi Mitra, ubli@usc.edu
Location: Hughes Aircraft Electrical Engineering Center (EEB) - -248
Audiences: Everyone Is Invited
Contact: Mayumi Thrasher
-
Diagnosis and Exploration of Massively Univariate Neuroimaging Data
Fri, Mar 03, 2006 @ 02:00 PM - 03:30 PM
Ming Hsieh Department of Electrical and Computer Engineering
Conferences, Lectures, & Seminars
Thomas E Nichols, PhDDept. Biostatistics,
University of Michigan, Ann ArborHost: Prof. Richard LeahyAbstract:For either fMRI or EEG/MEG, the massively univariate approach a linear model at each of, say, 100,000 spatial elements in the brain. The p-values computed at each voxel depend on assumptions on the data, and inferences can not be trusted unless these assumptions are checked. However, distributional assumptions are rarely checked in neuroimaging due to the sheer scale of the data. In lieu of examining 100,000 diagnostic plots, we propose a combination of statistical and graphical techniques to efficiently diagnose model fit. We create images of diagnostic statistics sensitive to typical model-violations, and time series of summaries that detect problem scans. Together with an interactive spatiotemporal viewer, we demonstrate how summaries can be used to swiftly find rare anomalies in millions of data elements We demonstrate the method on single-subject fMRI data as well as group-level fMRI data. One specific finding is that, while the popular SPM software assumes the temporal autocorrelation tis spatially homogeneous, we find dramatic variation of the autocorrelation strength over the brain, suggesting that fMRI data requires spatially-varying autocorrelation modeling.Biography:Thomas Nichols is an Assistant Professor of Biostatistics at the University of Michigan. He received his Ph.D. in statistics from Carnegie Mellon University in 2001 where he also trained in cognitive neuroscience at the Center for the Neural Basis of Cognition. He has been active in the field of functional neuroimaging since 1992 when he joined the University of Pittsburgh's Positron Emission Tomography (PET) Center as a programmer and statistician. Dr. Nichols' research focuses on modeling and inference of functional neuroimaging data, including PET and Functional Magnetic Resonance Imaging (fMRI). He has developed methods and software for: Nonparametric analysis of PET fMRI data, inference methods which account for the multiplicity of searching the brain for changes in activity (SnPM); diagnosis and exploration of massively univariate models fit on imaging data (SPMd); and high temporal resolution reconstruction methods for PET.
Location: Hughes Aircraft Electrical Engineering Center (EEB) - 248
Audiences: Everyone Is Invited
Contact: Regina Morton
