Illinois CS Students Win 2 of 13 Inaugural Google Fellowships

Two Illinois computer science students were recipients of the inaugural Google Fellowships. Ali Farhadi was selected as the winner of the Google Fellowship in Computer Vision Image Interpretation, and Eric Gilbert was chosen as the winner of the Google Fellowship in Social Computing.

The Google Fellowships were presented to exemplary PhD students in computer science or related research areas. 13 total fellowships were awarded.

"We have given these students unique fellowships to acknowledge their contributions to their areas of specialty and provide funding for their education and research," said Google in a statement. "We look forward to working closely with them as they continue to become leaders in their respective fields."

"Illinois computer science students continually impress with the excellence of their research," said Michael Heath, interim head of department and Fulton Watson Copp Chair in computer science. "These awards demonstrate that once again our students have not only a remarkable vision, but are addressing new challenges with outstanding results."

Ali Farhadi
Farhadi conducts his research on ways to improve object recognition and human activity recognition through transfer learning, scene discovery, and learning to describe objects with professor David Forsyth. His papers deal with two major themes: How can one recognize objects viewed from a new perspective using only examples from other perspectives? And how should one organize representations for object recognition?

To address the former, several of his papers deal with sign language. "For a variety of reasons, it is very useful to be able to spot words discriminatively in an ASL [American Sign Language] video (for example, one can use this capacity to align a video with a transcript)," said professor Forsyth. "Ali has shown that you could learn a word model from a frontal view of a dictionary signer (avatar), and then find the word being signed by a real person in a three-quarter view (a view from the side of the body). What is important to note here is that, in this experiment, the word spotter has never seen the word being made by a real person, and has never seen the word in a three-quarter view; conventional methods fail utterly under these circumstances."

Ali has several additional papers dealing with how to apply this principle of transfer learning to recognizing activities and recognizing objects.

His other work deals with one of the most frustrating features of modern object recognition theories: if you do not have a model of an object, you cannot say anything sensible about it. Ali has demonstrated how to change that, by showing how to represent objects in terms of their attributes. His work in this area attempts to automatically infer object attributes, such as "furry" or "boxy", from images. By doing so, one can describe unfamiliar objects when one encounters them, for example, saying that a monkey is "furry", "has a tail", "has a head", and so on. This work provides several new abilities, such as learning to recognize objects based purely on textual description and saying what is unusual about an object, while showing potential to yield more robust object detection algorithms that can learn from few examples.

Eric Gilbert
Eric's research addresses challenging and contemporary research problems in the rapidly emerging field of social computing. The purpose of social computing is to develop and apply principles of computing to create, sustain, and improve social interactions between people. Eric conducts his work with professor Karrie Karahalios.

In one research project, Eric was able to apply social capital theory in an analysis of thousands of MySpace profiles. Among many results, he was able to demonstrate that online relationships mimic those of the offline world, despite the opportunity to connect with millions of people online. The results advance our understanding of how people use social media sites and yield new insights for how to design such sites more effectively. This work was published in ACM CHI 2008 ("The Network and the Garden: An Empirical Analysis of Social Media in Rural Life"), the most competitive conference in the field of human-computer interaction.

A second research project demonstrated how to predict the social construct of tie strength. His approach brought together social science theory and advanced techniques in machine learning by analyzing traces of selected user interaction in a large data set collected from Facebook. The current "friendship" status on most social networking sits is binary - friend or not friend. This work predicts where a friend is between those two endpoints. He also showed how these results can be translated into design changes that can dramatically improve the overall user experience of the site. The work was recently accepted to ACM CHI 2009. Both of these publications received the Best Paper Award from ACM CHI. This award is highly selective as only 1% of all submissions receive this award each year.