Large Semantic Spaces that Support Multimedia Content in Internet-Scale Data Dissemination Models

Publish/Subscribe systems have been widely adopted as a loosely coupled communication paradigm for building large-scale distributed systems. Many real-world applications leverage the selective dissemination capabilities of the paradigm to exchange millions of events between geographically scattered entities. While large-scale publish/subscribe systems accommodate for the asynchronous and decoupled exchange of scalar events comprising textual and numerical values, they currently lack support for unstructured data types such as images and videos. In which case, publishers and subscribers have to be mediated by trained Classifiers and Object Detectors to facilitate content tagging for targeted data meta search-ability and dissemination. We argue that integrating such knowledge extractors into Publish/Subscribe Systems can allow for supporting multimedia content as native events but can also have many implications on the overall system performance which hinders its scalability. Whereas, leaving the models outside of the system and allowing event producers to manage their own knowledge extractors to label input content can introduce high semantic heterogeneity in large-scale environments. More precisely, the large semantic space of human level recognition creates an unbounded data space of object and image-level labels. In this thesis, we tackle these problems under two environments, namely broker-based publish/subscribe systems atop federated overlays networks and decentralized publish/subscribe systems atop peer-to-peer overlay networks. Due to the closed nature of federated overlays, we propose embedding resource efficient binary image classifiers and optimizing their execution to achieve scalability. However, in decentralised environments where the cooperation of participants is either not desirable or infeasible due to the dynamic nature of the network, we propose leaving such operators at the publisher end, and we tackle the semantic gap problem with an approximate event matching model based on distributional models of word meaning.

Team

Tarek Zaarour

Dr Edward Curry

Institution: NUI Galway

Funder

Relevant Publications