PONEDELJAK, 16. avgust 2010. u 11 sati:


Zoran Obradovic, Director, Center for Information Science and Technology, Temple University, Philadelphia, USA
STRUCTURED REGRESSION BY CONTINUOUS CONDITIONAL RANDOM FIELDS AND MULTIPLE NOISY ORACLES

We will first describe and evaluate our new kind of probabilistic model for structured regression called continuous conditional random fields (CCRF) that uses multiple non-structured predictors as its features. These features are constructed such that the result is a Gaussian predictor where learning is reduced to convex optimization leading to a global solution for a set of parameters while inference is conveniently conducted through matrix computation. Then, we will present a new Bayesian approach that learns a regression model from data with noisy labels obtained by multiple oracles. Our method provides a closed form solution for model parameters and in our experiments was consistently more accurate than a model trained with averaged estimates from multiple oracles as labels. Presented results are obtained in collaboration with Y. Guo, V. Ouzienko, V. Radosavljevic, K. Ristovski and S. Vucetic at Temple University. Both articles will appear at the Proc. 19th European Conf. on Artificial Intelligence, Lisbon, Portugal, Aug. 2010.

SPEAKER INFORMATION: Zoran Obradovic, professor of Computer and Information Sciences and the director of the Center for Information Science and Technology at Temple University in Philadelphia is an internationally recognized leader in data mining and bioinformatics. He has published about 220 articles addressing data mining challenges in health informatics, climate and ecological management, the social sciences, and other domains. Obradovic was the program chair at six, track chair at seven and program committee member at about 40 international conferences on data mining. He currently serves as an editorial board member at seven journals and is the executive editor at the journal on Statistical Analysis and Data Mining which is the official publication of the American Statistical Association (ASA).

Utorak, 31. 08. 2010. u 18h sala 2 SANU
Oskar Mencer, Department of Computing, Imperial College London
VERTICAL ACCELERATION: FROM ALGORITHMS TO LOGIC GATES

Abstract. Huge growth in the complexity of science over the past decade has driven the need for ever more computationally demanding mathematical models. This has in turn led to massive growth in data center compute capacity, and consequently power and cooling requirements. We report the real-world results from Maxeler Technologies to improve the price-performance and just raw speed for financial and seismic computations by using custom computing machines. The key to our acceleration approach is the vertical acceleration of key pieces of the algorithms, including the bit level, arithmetic level, architecture level, and algorithm level.

Speaker information: Prior to founding Maxeler Technologies, Oskar Mencer was Member of Technical Staff at Bell Labs, leading the effort on "Computing with FPGAs" within the Computing Sciences Center. He joined Bell Labs after receiving a PhD from Stanford University. Besides his work with Maxeler, Oskar is also affiliated with the Computing Department at Imperial College London and holds a Consulting Professor position at the Geophysics department at Stanford University. Oskar's experience includes start-ups and more established companies in Silicon Valley such as DIGITAL, Rockwell and Hitachi (Tokyo).