Posted by
When Chris J.C. Burges (opens in new tab) came to Microsoft Research in 2000, he knew he wanted to work on machine learning (opens in new tab) projects that would have a real impact on users.
Microsoft Research Podcast
Collaborators: Holoportation™ communication technology with Spencer Fowers and Kwame Darko
Spencer Fowers and Kwame Darko break down how the technology behind Holoportation and the telecommunication device being built around it brings patients and doctors together when being in the same room isn’t an easy option and discuss the potential impact of the work.
Burges definitely succeeded: He ended up being part of a team that created the basis for the ranking system that is still used in Microsoft’s Bing (opens in new tab) search engine today.
At next week’s International Conference on Machine Learning (opens in new tab), Burges, a research manager and principal researcher in Microsoft Research’s Machine Learning Intelligence Group (opens in new tab), and his co-authors will receive the Test of Time Award (opens in new tab) for the 2005 paper (opens in new tab) that showed how that system works.
The system, called RankNet, was a breakthrough because it was much faster and more accurate than the previous system. Burges said it was able to make as much progress training the search engine system to rank results in one day, using one PC, as a previous system had done in several days using a cluster of computers.
RankNet also relies on neural networks (opens in new tab). These are computer systems, loosely modeled after the human brain, that can be trained to perform desired tasks based on data labeled by humans. At the time, Burges said he believes they were the only researchers using that technology for a search engine ranking system.
In the last few years, the use of neural networks has exploded, with researchers using them to make great strides in everything from real-time translation (opens in new tab) to image captioning (opens in new tab). The program co-chairs lauded Burges and his team for showing the power of these networks way back in in 2005, long before this most recent renaissance.
Burges said he was drawn to search engine ranking because it was a hot, highly competitive field in which both researchers and technology companies were competing to be the best.
“We knew there was a significant opportunity for having impact,” he said.
These days, Burges is working on a system that aims to teach machines (opens in new tab) to read and comprehend text, and to be able to answer questions about it.
It’s an ambitious project that, if successful, could have profound implications for artificial intelligence, or the development of systems that can see, hear and understand.
Burges said one thing he’s learned in his career is that it’s worth taking on longshot projects that you are passionate about.
“At some point you should just say, ‘What do I really want to accomplish?’” he said. “And then you should just do it.”
A number of Microsoft researchers are presenting other papers at the machine learning conference. They include:
- Approval Voting and Incentives in Crowdsourcing (opens in new tab) (611 KB .pdf)
- A Lower Bound for the Optimization of Finite Sums (opens in new tab) (232 KB .pdf)
- Surrogate Functions for Maximizing Precision at the Top (opens in new tab) (435 KB .pdf)
- Optimizing Non-decomposable Performance Measures: A Tale of Two Classes (opens in new tab) (469 KB .pdf)
- Classification with Low Rank and Missing Data (opens in new tab) (363 KB .pdf)
- Stochastic Primal-Dual Coordinate Method for Regularized Empirical Risk Minimization (opens in new tab) (426 KB .pdf)
- DiSCO: Distributed Optimization for Self-Concordant Empirical Loss (opens in new tab) (395 KB .pdf)
- A Linear Dynamical System Model for Text (opens in new tab) (335 KB .pdf)
- Multi-Task Learning for Subspace Segmentation (opens in new tab) (163 KB .pdf)
- On Greedy Maximization of Entropy (opens in new tab) (323 KB .pdf)
- Bimodal Modelling of Source Code and Natural Language (opens in new tab) (448 KB .pdf)
- Intersecting Faces: Non-negative Matrix Factorization With New Guarantees (opens in new tab) (452 KB .pdf)
- Un-regularizing: approximate proximal point and faster stochastic algorithms for empirical risk minimization (opens in new tab) (397 KB .pdf)
- Pushing the Limits of Affine Rank Minimization by Adapting Probabilistic PCA (opens in new tab) (374 KB .pdf)
Allison Linn is a senior writer at Microsoft Research. Follow Allison on Twitter (opens in new tab).
Artificial intelligence