SOCS Graduate Seminar Series Seminar Schedule

Does your work involve matrix operations with large dense matrices? Not satisfied with speed you are getting on CPUs? Then you may be aware that modern GPUs provide considerably better performance on many types of matrix computations than most CPUs. However, so far you had to choose vendor-specific APIs (eg: Nvidia CUBLAS only works on Nvidia) making your code non-portable. I present a high-performance portable BLAS library that runs on many GPUs including AMD, Nvidia and Intel GPUs. I will present initial performance results, and discuss future plans.

We describe a novel online topic modeling framework to compute a low dimension descriptor of visual observations made by a mobile robot, which is sensitive to the structural and thematic changes in the environment. Our approach is designed to run in realtime, and is suitable for long term execution on a robotic platform. Using this image descriptor we build online anytime summaries consisting of surprising observations experienced by a robot thus far. The observations in the summary are chosen such that they cover the set of all observations in topic space, while minimizing the cover radius. Like almost any summarization method, the technique is meant to produce data for human consumption. Thus, we assess our approach on 307 human subjects and compare it to the classic bag-of-words description based summaries, and find it superior.

The transmission of information through noisy communication channels is a fundamental problem of theoretical and practical importance. Given a probabilistic model of the noise in a communication channel, we can use Shannon's information theoretic techniques to calculate the maximum possible communication rates for that channel. Recently there has been interest in studying quantum channel models, which offer more accurate models for the encoding, transmission and decoding of the information. These models are relevant for communication systems in which the information carriers are quantum systems, like for example in the case of optical communication. This presentation will give an intuitive explanation of the Holevo-Schumacher-Westmoreland Theorem, which gives a precise characterization of the information carrying capacity of quantum channels. In closing we will discuss briefly the generalization of the HSW Theorem to quantum channels with many senders and/or many receivers (network information theory).

Ivan Savov is a graduate student in the School of Computer Science at McGill Univeristy in Montreal Canada. Previously he received a B.Eng. degree in electrical engineering in 2005 and a M.Sc. degree in Physics in 2008, both from McGill University. His current research interest include network information theory, error correcting codes, quantum information theory and machine learning.

I have recently taken part in the Nokia Mobile Data Challenge (MDC). Nokia has provided over a year's worth of data collected from the mobile phones of 38 people in Lausanne, Switzerland. The data include wifi, gsm, and gps signal measurements, accelerometer measurements, call logs, and application usage. I will present a set of tools that I have developed to visualize and analyse this data, as well as some preliminary results from our paper "Generating storylines from mobile phone data" (submitted to the Nokia MDC Workshop), in which we proposed a method for translating from mobile phone sensor data to English-language descriptions of events.