The PASCAL network of excellence is coordinated by John Shawe-Taylor (Scientific Coordinator) at University College London, U.K. and Steve Gunn (Operational Coordinator) at the University of Southampton, U.K.. In total there are 68 sites in the network.

Gender Statistics of Researchers in PASCAL

Leveraging Complex Prior Knowledge for Learning Thematic Programme
1 March – 30 September 2008

Traditionally machine learning has focused mainly on constructing models in a data driven manner. Clearly, in practise, if we can incorporate domain knowledge with our learning we should be able to obtain improved performance. This type of knowledge is particularly important in application domains where data availability is sparse in the context of the complexity of the required model. In this thematic programme we will highlight and drive forward approaches to incorporating prior knowledge in the application domain. We are interested in all approaches to incorporating this prior knowledge and any application area. Already some subthemes (and application areas) are emerging within the programme for example: knowledge encoded in graph structures (applications in language and computational biology), knowledge encoded in ordinary and stochastic differential equations (applications in climate and systems biology) and knowledge encoded as probabilities (applications in language). The Thematic Programme will run between March and September 2008 with a potential extension period to March 2009.

 Muilti-Component Learning Thematic Programme
1 October 2008 – 28 May 2010

Computer systems seldom operate in isolation and the outcome of learning tasks on one component may affect a related task on another. For example learning how best to redirect network traffic will once implemented affect the solution that should be adopted at an adjacent node. Cognitive systems composed of multiple agents are another example in which different components may be adapting their behaviour to achieve certain goals, the effects of which will influence the operating environment of other components. The design and analysis of systems involving interacting learning systems is still in its infancy, particularly when we consider theoretical analysis that can be used to guide their design, and if we include self-organisation as a design principle. A related set of challenges arise when we consider integrating information from diverse sources as for example in distributed sensor networks. Once again learning must be used to decide how to filter the data to ensure the network can provide informed responses to a range of different queries. Learning at one node of the network will influence the optimisations at other nodes. The key objective that can enable solutions in all of these applications is to build a well-founded theoretical framework analysing learning in a game theoretic setting. The learning approach can deliver the flexibility, robustness and scalability that are properties required for many applications of cognitive systems, for example in robotics. Such a framework can then provide the criteria that can be used to design and optimise multicomponent systems for a wide range of applications.

Partial or Delayed Feedback Thematic Programme
1 June 2009 – 31 December 2010

This Thematic Programme will foster research on learning models and algorithms when – in contrast to supervised learning – information about the correct predictions are not immediately available to the learner. The assumption of full information about a training instance is often unrealistic and in many applications the learner must deal with limited feedback. Although some aspects of learning with limited feedback have already been thoroughly analyzed (e.g., multi-armed bandit problems), many problems are still open.

Among others the following topics are relevant for this Thematic Programme:

• Reinforcement learning as a model of delayed feedback, where the utility of predictions/actions might be revealed only after a number of further predictions.
• Variants of the bandit problem as models of partial feedback, where only the utility of the learner’s predictions is available but not the utility of possible alternative predictions.
• Models of indirect feedback, where neither the true outcome nor the utility of the prediction is observed, but only an indirect feedback loosely related to the prediction.
• In general, the exploration-exploitation trade-off in learning models.
• Semi-supervised and active learning.

Cognitive Architecture & Representation Thematic Programme
1 October 2009 – 28 February 2011