Research

The core research question of DIRAC concerns how to identify samples of data that do not fit into any of the classes seen during training, while at the same time constituting a valid object in the sense that the new sample is not merely a noisy measurement.

The notion of a hierarchy of processing and classification levels has been identified as essential. Such a hierarchy is defined in terms of two views, parts-whole and class-memberhship partial orders. A "rare" event is detected when classifiers at different levels of the hierarchy give different -incongruous- results. The paper and presentation below give a detailed exposition of the theory of incongruous event detection:

• Weinshall D., Hermansky H., Zweig A., Luo J., Jimison H., Ohl F., Pavel M. "Beyond Novelty Detection: Incongruent Events, when General and Specific Classifiers Disagree", in 'Advances in Neural Information Processing Systems (NIPS)', Vancouver, December 2008.
• paper
• presentation
  

Current work targets the definition of a partial order representation for which we would be able to learn its structure given labeled training data. This structure will enable the recognition of incongruent events following our previous research results. Following our unifying approach for class-membership and part-membership hierarchies, we suggest a single partial order structure, representing simultaneously both hierarchies. We view the task of grouping classes together and the task of finding co-occurrences of parts belonging to different class as dual problems. Given a representation of an object class as a set of parts, a grouping of several classes into a single more abstract class defines a co-occurring set of parts shared by these classes. On the other hand a set of co-occurring parts defines a grouping of one or more classes sharing these parts. The connection between these two concepts is achieved by the notion of partial order. A hierarchy of classes defines a partial order among the classes as do containment relations among set of properties define partial order over these sets.

Given such a partial order representation, a sample from a new class represented as a novel combination of parts would be accepted be all general level classes which are  represented by known groups of parts which appear as subgroups of the set of parts representing the sample of the novel class. As no combination of such parts has been seen before, there will exist no specific level accepting this sample, thus the model will contain accepting nodes at higher levels in the partial order and lower levels will reject.