The Curse of Dimensionality
Yoshua Bengio wants to beat the curse of dimensionality. To succeed, he's going to have to teach a machine to overcome this problem-solving quagmire.
Bengio is a leading international expert in the fields of neural networks and machine learning. The applications of these computer-based techniques apply to a broad range of cutting-edge topics-from speech-recognition software to hastening the pharmaceutical development process.
The current research of this new Canada Research Chair at the Université de Montréal focuses on "learning algorithms," which are problem-solving methods that computers use to actively "learn" from received data.
Along with a research team of a dozen students and postdoctoral fellows, Bengio will focus on two central problems. First, developing learning algorithms and techniques to deal with sets of data that contain numerous variables. And second, to create better learning algorithms for very large sets of data, with tens or hundreds of millions of pieces of information.
The curse of dimensionality faces Bengio in those cases which involve numerous variables. The challenge is that in developing probability models, the number of possible combinations increases exponentially with the number of variables. It's a major hurdle for such applications as speech recognition, translation and information retrieval (e.g., Web search engines) software-cases in which the possible combinations of words to make up a sentence can run into multiples of a billion. Professor Bengio will design improved machine-learning techniques for these high-dimensional data sets. His research and development in this field will clearly have a significant impact on industries that rely on machine learning and complex computer software.
This computer scientist will also create improved learning algorithms for "mining" large amounts of data. Professor Bengio will experiment with a variety of techniques to increase the learning algorithm speed for these large data sets. To accomplish this, he is working with corporate partners, including insurance companies. One of the goals here is to develop improved ways to determine the risk/premium ratio when setting insurance rates.