The DAE Platform: A Next-Generation Framework for Experimental Pattern Recognition Research
Brown Bag Lecture by Dr. Bart Lamiroy | 7/26/2011 11AM-12PM | 7th Floor Conference Room, Bldg 38A
Abstract: The development of effective algorithms for information retrieval or pattern recognition tasks requires not only validated data for testing, but also an environment that allows comparisons with competing algorithms, capable of guaranteeing reproducible and verifiable experiments and which enables keeping a history of such experimental activities. DAE is a platform for this purpose.
The DAE project started off as a DARPA overseen project at Lehigh University in 2010. Its initial aim was to provide an international resource for document analysis research. It has rapidly grown far beyond its initial scope. This talk will give both an overview of what the project has achieved from a technical point of view, but, more importantly, it will give an account of the great research opportunities it has helped shape and the innovating light it has shed on experimental research in general, and both pattern recognition and machine perception in particular. We will also discuss the role of NLM’s validated data resources in this enterprise.
Indeed, “providing resources” to an international research community, be it Document Analysis or any other, requires us to assess what kind of resources to provide, why they are interesting or important, why they are currently missing in the common research play field and whether they would actually impact research quality when provided. Answering these questions requires some introspection into our own and our communities’ research practices, identifying invariant patterns, pinpointing and analyzing inefficiencies or sub-par best practices and conceiving quick-win resources that would improve the quality of the related research. During the talk we shall examine all those constraints and present the technical choices and architecture that was retained for the DAE platform. The resulting tool will be demonstrated, and the lessons learned from its implementation will then be applied to a more broader scope in a more long term perspective. What if this tool could leverage certifiable experimental research in the long term ? How would its current implementation help to achieve it ? What should be further developed ? What scientific challenges does this raise ? What actual achieved milestones encourage us to proceed ?
Bio: Bart Lamiroy is a permanent faculty member at Nancy Université – Institut National Polytechnique de Lorraine, in Nancy, France, and member of the associated LORIA research lab since 2000. He was a visiting scientist at Lehigh University from January 2010 through July 2011 and is currently a visiting scientist at NIH/NLM. He has a broad experience in Machine Perception. Over the years, his research topics have ranged from Content Based Image Retrieval over Visual Servoing to Document Analysis.
From 2007 to 2009, he was head of the Computer Science and IT Department at the École des Mines de Nancy, France, and has been a permanent faculty member there since 2000. Before that he was a research contractor at INRIA, after having obtained his Ph.D. in computer vision at the Institut National Polytechnique de Grenoble, France in 1998. He received his bachelor’s degree in applied mathematics in 1993. He also serves on the International Association for Pattern Recognition TC-10 Committee as Dataset Curator, and on the Publicity and Publications Committee.