Workshop on December 12-13, 2008, Whistler, BC, Canada
- Slides and recordings of the workshop talks are now available from videolectures.net
Kevin Kelly, former editor of Wired magazine, stated in an inspiring blog article that the internet has reached the stage where it can be looked upon as a gigantic information copying and distribution mechanism. But when the problem of distributing and copying information is essentially solved, where do we go next?
It is in particular the networked structure of devices and systems that generates a host of interesting problems and opportunities. So what are the values that can be derived from this mesh, values that go beyond the fact of copying and distributing information? In his article, Kelly identified a number of such values, a few of which have immediate relevancy for the ML community.
- Interpretation: Copying information can be achieved at a cost of zero, but how do we make sense of it?
- Immediacy: How can, for example, a piece of news be delivered immediately (at the moment it is produced) to an interested reader?
- Findability: Or: An unfound masterpiece is worthless.
- Personalization: Via an ongoing conversation between producer and user, tailored versions of an otherwise generic piece of work can be generated.
- Privacy: Access to data needs to be restricted to a necessary minimum, to protect identities
In order to implement these values for the human that is linked into the mesh of information and devices, research in a number of directions needs to be carried out:
- Machine learning and probabilistic modeling: Recommendation systems and knowledge extraction are two immediate applications. Current work needs to be improved in a number of aspects, including large scale inference, modeling languages, and efficient decision making.
- Game theory and mechanism design: When a large number of contributors are involved in solving a particular task, how can we set up the tasks and the incentive structure such that "the crowd" does indeed achieve the desired goal? Game theory can provide answers to how the interaction of large groups of people can be modeled, whereas mechanism design is concerned with the design of rules that lead to desired behavior. Research is required in particular for solving very large games, and for mechanism design under uncertainty.
- Knowledge representation and reasoning: Large parts of the current internet data are stored in an unstructured way, making linking and evaluating knowledge a complex problem. Semantic web solves parts of the representation problem, but the question of reasoning is still largely unsolved. Also, there is a tradeoff to be made between efficiency of reasoning and power of the representation. Reasoning under uncertainty is a further challenge here.
- Social networks and collective intelligence: The web in its current form can also be viewed in terms of its interaction graph. Who is sending an email to whom? Who is reading whose blog entries? How does information flow in these networks? To answer these questions, the networks need to be analyzed, modeled, and made amenable to reasoning.
- Privacy preserving learning: For the above questions to be solved, it is often necessary to share personal information of users - which, on the other hand, user may not want due to the loss of their privacy. What can be learned, and how can be learned, whilest only revealing a minimal set of information, or information that does not make users individually identifiable?
Making progress in these areas will require a large deal of research carried out across the typical boundaries. We wish to make this workshop as much interdisciplinary as possible, in that it addresses all of the above topics, and highlights the connections. We tried to choose speakers (and will select contributions to the workshop) that work on the connections between two or even more of these topics. We aim at addressing topics both from an academic and from an application point of view.
Ideally, the outcome of the workshop would be to
- Provide a new research agenda that links the above topics
- Identify a set of reachable next steps
- Provide the research communities with a realistic working environment, that is, datasets.
A number of researchers in different research communities are working on problems like machine learning for web search, on recommender systems, on game theory, or on generic methods for reasoning. We believe that there is significant value in providing a forum for discussions amongst these researcher, to identify chances for collaborations and exchanging ideas that will help make the developed methods applicable on web scale.