Two electrical applied scientific discipline professors from the Jacobs School of Engineering, Gert Lanckriet and Nuno Vasconcelos, have won spots inwards the 2013 Yahoo! Faculty Research in addition to Engagement Program (FREP).
Check out the total listing of awardees at the Yahoo! Labs Faculty Research in addition to Engagement Program website.
This programme funds assistance academics across the globe collaborate with Yahoo! query scientists on new, exciting cyberspace query studies in addition to experiments.
Lanckriet in addition to Vasconcelos, from the Department of Electrical in addition to Computer Engineering at UC San Diego, are 2 of 27 recipients from 24 universities inwards vii countries were selected based on their areas of scientific discipline in addition to involvement to Yahoo!.
A summary of Lanckriet’s projection is below.
Machine Learning with Highly Unstructured Multimedia DataAn increasingly of import aspect of information inwards today’s footing is that it is highly unstructured in addition to heterogeneous. For example, a search for “miss footing 2013” yields fundamentally dissimilar modes of information close the theme via intelligence articles, blogs, videos, pictures etc. While at that spot are roughly commonalities inwards the information provided inwards these modalities (i.e., they are close the beauty pageant), the nature of information differs widely alongside the modalities.
Videos in addition to pictures are innately visual inwards nature, in addition to differ from intelligence articles in addition to blogs which are textual inwards nature. Even inside the textual sources, at that spot be differences: intelligence articles are to a greater extent than factual inwards nature, spell blogs are opinionated. In multimedia applications, such information is oftentimes available inwards a diversity of unstructured formats: blogs may comprise i or to a greater extent than pictures, or none, embedded videos, well tracks, and/or tweets, or not, etc.
Ideally, an algorithm to brand inferences from the information (e.g., whether or non the judges were fair inwards deciding the winner) would hold out given access to all available information sources, since each origin may convey roughly unique information to the table. Machine learning with such unstructured, heterogeneous information is challenging in addition to an active theme of research. Existing approaches assume a rigidly structured input format. They cannot adjust missing modalities (incomplete information) or modalities with multiple instances (over-complete information). In this project, nosotros retrieve to pattern in addition to railroad train principled approaches to learning from heterogeneous, unstructured multimedia data.
One of the problems that nosotros are interested inwards is creative someone recommendation. Current approaches rely on either collaborative filtering methods (relying on the wisdom of the crowd to arrive at upward one's heed if 2 artists are similar), or content based methods (two artists are considered like if the well content of their songs are similar) to arrive at upward one's heed if 2 artists are similar. We are interested inwards developing a multimodal creative someone similarity, based on heterogeneous unstructured social in addition to multimedia information that relates to artists. This tin include band videos in addition to pictures (which implicitly supply information close the “look-and-feel” of an artist), tweets, reviews, in addition to blogs close the artists’ music (providing opinions in addition to information close artists), lyrics, their social network of listeners, etc. We believe that combining all this information through a to a greater extent than accurate multimodal similarity mensurate volition ameliorate recommendation compared to using content or collaborative filtering only.
Sumber http://jacobsschoolofengineering.blogspot.com/