A large collection is valuable only if users can find what they want, when they want it. The key issue is describing and recognising media content in a fashion that is consistent with user needs. As a result, a negotiation between user needs and technological capacity is inevitable, and a good knowledge of likely user requirements can make the difference between success and failure.

Users are often interested in semantic categorisation, which is seldom available; the trick is to construct recognition processes that usefully mimic semantic categories. Recognition is poorly understood, but the most useful mathematical ideas appear to be statistical in nature, including classification, statistical learning theory, and probabilistic inference; often "natural" recognition strategies (such as Bayesian inference) are computationally infeasible, and there is a dearth of results that link statistical accuracy with computational cost in useful ways.

Search is made easier by summarizing content. This trick works at a variety of levels: for a collection, one might use a very small subset of images (or videos, or audio tracks, or...) to indicate both the type of content and the rough distribution of types; similarly, representing a video by a few well chosen frames can make interactive browsing fast and efficient. Summarization draws mainly on transform and coding theory.

Current search tools are weak. Browsing tools can make weak search tools very useful, by making it easier to navigate from the response of a search tool to the right place in a collection. Good browsing tools are fast and responsive, and use the display to guide the user to an understanding of the local structure of the collection. Isometric embeddings of graphs and Haar measure are mathematical tools that have proven useful here.

The workshop will:

• acquaint researchers with the state of the art of knowledge in user needs and the technology available to meet these needs
  
  
• drive theoretical research in areas such as learning, classification and inference by describing the technology needs of system builders
  
  
• build collaborations between researchers with appropriate theoretical backgrounds and system builders.

Appropriate attendees are: mathematicians and statisticians interested in classification, learning and inference, in transforms or coding theory or in applications of geometry; multimedia practitioners interested in retrieval of various forms of content, including documents, images, video and music; and multimedia practitioners interested in managing, browsing or visualising large collections of data.

WORKSHOP SCHEDULE

LIST OF CONFIRMED PARTICIPANTS
(in addition to postdocs and long-term participants)

2000-2001 Program: Mathematics in Multimedia

Talk Abstracts   Material from Talks