Project Manager : Hervé GLOTIN
Keywords : Information Dynamics, Integration Dynamics, Content Based Information Retrieval, Prediction, Machine Learning, XML Extension, Data Bases, Signal Processing, Artificial Perceptual Intelligence, Neurophysiology, Deep Learning, Semantic Gap, Cognition
Application Keywords : Image, Video, Speech, Web Ressources, RSS, Bioacoustics, Broadcast News, Named Entity Recognition, Cetacean
Robust multimodal information analysis is hindered by the limitations of current computer systems given the increasing demands for computational efficiency and storage.
The DYNI project offers a theoretical framework for a scaled analysis of these high-dimensional spaces. Our approach aims to unify different levels of abstraction and dynamic processing as they relate to diverse data such as speech, video, image, collections of texts and other. The strength of the approach lies in its concentration of interdisciplinary expertise in signal processing, automatic transcription, and design of flexible models for information retrieval, query and data transformation to unified standards like XML.
Specifically, DYNI addresses the challenge of representing data through their abstract content allowing the creation of generic solutions, independent of the data that can lead to the bridging of the semantic gap.
DYNI is formalized based on theories of signal processing, stochastic
estimation, machine learning, dynamic filtering and on multidimensional databases, languages and manipulations. It includes two research axes that mutually reinforce each other: the integration of temporal data, and the dynamic integration of concepts in a more cognitive framework.
The first research Axis concerns the integration of temporal data, or more generally the integration of forward flow such as speech, video, biological data, web feeds, etc. For example, multimodal indexing by the content, bio-acoustics indexing, or filtering data at different levels of representation leading to manipulation of XML data providing high-level entities.
Projects in this research area are part of the French Reseach Agency (ANR) as COGNILEGO (with pole OPTITEX), ORIGAMI, ROSE and OTIM. Other contracts exist with the Nation Armement Direction (DGA), and industrials as DCNS, A2IA, and Chrisar Software or Cesgima PME within
programs supported by the PACA region (pole SEA).
We regularly evaluate our models through international campaigns in video: TRECVID 2008-09-10-11 (Nat. Instit. Stand. Of Tech) within the national IRIM consortium, or ImageCLEF ROBOT 2009, MEDIAEVAL 2011. Or also on acoustic international challenges : ESTER (Evaluation System Transcript Enriched radio programs - DGA IRISA), detection and tracking of cetaceans in passive acoustics (DLMMAP 2005, 2007, 2009 and 20011 - Filing international patent).
The second Axis focuses on the dynamic integration of concepts,
especially for static data. It involves modeling the interactions
between different levels of representation and processing. We believe
that this analysis leads to intrinsic dynamics of information systems,
multi-scale processing and semantic analysis.
These researches are conducted through ANR project for modeling the
learning of reading (ANCL project), and automatic annotation of images
on the web (AVEIR project). Moreover, a thesis on the issue of scaled
and approximate learning is running in collaboration with the LIF and
inside the FRIIAM Federation (Fédération Informatique et Interactions
Evaluation in this framework are conducted within participations in
the competition ImageCLEF Photo-Annotation in 2008 and 2009 (top 10),
the task of Diversity ImageCLEF 2008, Imageval (led by CEA) (top3).
A more recent project is concerning sportive strategy analysis.
In sum, the first research Axis of DYNI will lead to propositions of
new standard methods for the integration and flexible manipulation of
dynamic multimodal information. The same methods will be extended to
the concept of dynamic integration (possibly inherently static),
reflecting the second research Axis.
These activities are unified through common signal and stochastic
processes, and an information system approach. We think that this dual
strategy offers a strong framework for bridging the semantic gap and
developing information systems with high power generic expression