| Speaker | Abstract |
|---|---|
27-01-2010 |
High resolution sampling of smooth signals using low-precision quantizersThe problem of sampling a bandlimited or smooth non-bandlimited signal with a bounded dynamic range with low-precision quantizers is addressed. A sampling unit, having access to the samples gathered by low-precision quantizers, is required to reconstruct the signal to maximum accuracy. Bandlimited signals:The feasibility of having a flexible trade-off between the oversampling rate and the quantization-precision, while achieving the best known exponential accuracy in the number of bits per Nyquist-interval is demonstrated. This exposes an underlying "conservation of bits" principle. Extensions of these results are established in an almost sure sense for amplitude-limited stationary (stochastic) bandlimited fields.Non-bandlimited signals:If sampling precedes anti-alias filtering operation, then aliasing error will be present during sampling of non-bandlimited signals. This problem is motivated from distributed sampling constraints, where an anti-aliasing filter cannot be used prior to sampling a spatial field. A framework, which accounts for aliasing as well as quantization errors, is developed for the sampling of smooth non-bandlimited signals having finite first absolute spectral moment. Using this framework, upper bounds on the maximum pointwise reconstruction error will be derived as a function of bit-rate (per meter) spent in recording the signal. Similar to the case of bandlimited signals, a flexible trade-off between oversampling rate and quantizer-precision is obtained. |
05-02-2010, 15:30pm |
A simple necessary and sufficient condition for the double unicast problemWe consider a directed acyclic network where there are two source-terminal pairs and the terminals need to receive the symbols generated at the respective sources. Each source independently generates one symbol from a given alphabet in an i.i.d. manner per unit time. Each edge in the network is error-free, delay-free, and can carry one symbol from the alphabet in unit time. We give a simple necessary and sufficient condition for being able to simultaneously satisfy the unicast requirements of the two source-terminal pairs using network coding. Extensions and other applications will also be presented. |
19-02-2010, 2:30pm |
Human Activity Recognition at Mid and Near Range from a Single Video StreamRam Nevatia is a professor in the Computer Science Department, University of Sothern California |
19-02-2010, 3:30pm |
Two-Way Transmission Capacity of Wireless ad-Hoc NetworksThe transmission capacity of an ad-hoc network is the maximum density of active transmitters in an unit area, given an outage constraint at each receiver for a fixed rate of transmission. Most prior work on finding the transmission capacity of ad-hoc networks, has focused only on one-way communication where a source communicates with destination and no data is sent from destination to the source. In practice, however, two-way or bidirectional data transmission is required to support control functions like packet acknowledgements and channel feedback. This talk extends the concept of transmission capacity to two-way wireless ad-hoc networks, by incorporating the concept of a two-way outage with different rate requirements in both directions. |
26-02-2010 |
Automatic Melody ExtractionThe melody of a song is an important attribute in music information retrieval applications, such as query-by-humming, and also in musicology and singing pedagogy. Automatic extraction of the melody from polyphonic vocal music is a challenging problem for which no general solutions currently exist. In this talk, we present a melody extraction system, which utilizes a spectral harmonic-matching pitch detection algorithm followed by a dynamic programming-based optimal path finding technique that tracks the voice-pitch within certain melodic smoothness constraints. Extenstions and enhancementst will also be presented. |
TBA |
Optimal Timer Schemes for Distributed SelectionMany wireless communication systems envisage the use of a selection mechanism to discover or select the most suitable candidate, from a set of many available candidates. One such example is cooperative systems that has to select the best available relay to exploit spatial diversity. Here, system consists of n wireless nodes. Each node possesses a suitability metric that is known only to that specific node. Our aim is to make a central node determine which node has the highest metric. A simple and effective technique for this is to use timer schemes. A node transmits a packet when its timer expires, where timer value is a monotone non-increasing function of its metric. The best node is selected successfully if no other node’s timer expires within a vulnerability window after its timer expiry, and so long as the sink can hear the available nodes. In this talk, we will present optimal timer schemes that (i) maximizes the probability of success or (ii) minimizes the average selection time subject to a constraint on the probability of success. |
19-03-2010 |
"What's Popular Amongst Your Friends?"Recommendation systems, like those employed by Netflix and Amazon, suggest relevant content to potential buyers. Given the lack of good statistical models and the massive datasets, the dominant research focus has been on proposing algorithms and testing their scalability and performance on specific data sets. Recently, however, there is emerging interest in provably good techniques. For example, several authors have taken a compressed sensing view of this problem, and derived bounds on the least number of samples required for matrix completion. In earlier work (with S. Aditya and B. Dey) we proposed a new viewpoint motivated by channel coding. In this talk, we further exploit this viewpoint to analyze a simple algorithm motivated by practice. The algorithm considered makes recommendations based on popularity amongst "like-minded" users. The algorithm has competitive performance on the Movielens and Netflix data sets, and for a mathematical model, we provide a detailed bit-error-rate (BER) analysis. In particular, the analysis reveals three distinct performance regimes, and gives insight into the performance of the scheme on real datasets. |
TBA |
TBA |
TBA |
TBA |
TBA |
TBA |
Send suggestions and/or comments to: bsraj att ee .