SEATTLE, Aug. 8 – The vast majority of Microsoft research–included in the firm’s $5.3 billion R&D budget for FY 2002–is for AI-related projects, Microsoft Chairman Bill Gates said, speaking at the International Joint Conference on Artificial Intelligence.
Microsoft’s research is focused on traditional AI areas, such as decision-making, learning, language, and speech recognition. “We are putting our money where our beliefs are: that these things will become real and allow us to build far, far better software products than we have today … for many hundreds of million if not billions of people who will be using and taking advantage of every day.”

Microsoft researchers showed several examples of current projects:

Distributed speech recognition with natural language processing to automatically fill in a form on a handheld wireless device, using a remote PC for processing.

Data mining tools: a dependency network, using a graphical display to see patterns of TV show viewing and Web Canvas, a data mining tool for analyzing visitor behavior on Web sites.

Search tools: SWISH (Searching With Information Structured Hierarchically) automated clustering of Web search results to form easy-to-browse natural categories to simplify searching. The tool uses machine learning techniques to build statistical models and automatically create directories.

Automated question answering, using information from the Web

Automated sensing and reasoning about context and interests to classify incoming messages by priority and relevance. Each incoming email message gets a dollar value and priority based on what the recipient is doing at that moment.

Visual analysis to determine recipient activity

Real-time spectral analysis to localize speakers in an office

“We’re proceeding full-speed ahead,” said Gates. “We look forward to delivering your work to hundreds of millions of users.” However, in a Q&A, Gates admitted some AI failures so far in “social user interfaces,” such as Microsoft Bob and the much-criticized Clippy, which he described as “too obnoxious.”

“Unless there’s a deep model behind those anthropomorphic interfaces–until you cross over some threshold–it’s like speech recognition: it demos well and in practice, it’s kind of irritating. In think when you stand in front of your TV set in say, five years, we will have a rich-enough model. So we’re unabashed in terms of the approach, despite admitting our complete failures today.”