Transcript Slides

Challenges in
Information Fusion Technology Capabilities
for Modern
Intelligence and Security Problems
Speaker: Prof. Sten F. Andler
Director, Infofusion Research Program
University of Skövde, Skövde, Sweden (*)
Author: Dr. James Llinas
Center for Multisource Information Fusion
University at Buffalo, Buffalo, New York, USA
[email protected]
(*)
Key Information Fusion Challenges
Driven by Operational Problems and Modern IT
• Heterogeneity of Data, Information
• Common Referencing and Data Association
Impacts
• Dealing with Semantics
• The Entry of Graphical Methods
• Architecting Systems and Analytic Frameworks
Heterogeneity of Data/Information
Heterogeneity from modern IT capabilities/problems and networked systems
Lack of reliable a priori knowledge to support dynamic deductively-based reasoning
 “Weak Knowledge” problems
• Observational
– “Hard” Sensor Data and “Soft” linguistic/reported/unstructured Data
• Open-source & Social Media
– Issues: Mostly in linguistic form; Trust, Volume, Formats, Modalities
• Contextual differences
– Issues: Format, Middleware reqmt, dynamics, relevance
• Ontological differences
– Issues: Multiple-ontology cases, semantics, dynamics, relevance
• Learned knowledge
– Issues: integrating inductive and other inferencing procedures
Some Impacts due to Data Heterogeneity
• Soft (linguistic) data -- New preprocessing Front Ends: requirement for
semantically robust Text Extraction/NLP processes
– Marginally available today
– If not extracted, properly labeled entities never enter the Fusion process
– If not tagged with some level of (reliable) uncertainty/confidence, entity
uncertainty not considered
• Confounds both Common Referencing and Data Association
• Exploiting Contextual Data requires Middleware to condition data in a
form useable by Fusion process (native form-to-useable form)
– Can also require hybrid algorithms, eg context-aided Kalman Filter designs
• In networked systems, there can be multiple Ontological versions being
used
– Creates a need for ontological normalization (Common Referencing
function)
– Also impacts Data Association; inconsistent nomenclature will prevent
feasible associations
• Information learned in real-time creates a Level 4 Knowledge
Management functional requirement, and real-time adaptation that
can include dealing with out-of-sequence evidence (retrospective
adaptation)
Some further Impacts regarding
Common Referencing and Data Association
• Common Referencing
– Temporal alignment within streaming Soft data feeds is
challenging
• Dealing with linguistic tense: past/present/future
– Impacts correct Temporal Reasoning
» Creates a need for agile Temporal Reasoning
– Networked environments open the possibility for
inconsistent forms of uncertainty representation
• Creates a need for uncertainty transforms, normalization methods
• Data Association
– Major impact due to Soft (linguistic) data and availability of
Relational links
• Association now of higher dimension: Entities/attributes and interentity Relations — becomes a Graph Association problem
• New scoring functions required; eg Relational similarity
Representative Impacts regarding
Common Referencing and Data Association, cont.
G. Tauer, R. Nagi, M. Sudit, The graph association problem: Mathematical models and a
lagrangian heuristic, Naval Research Logistics (NRL) Volume 60, Issue 3, pages 251–268, April 2013
Representative Impacts regarding
Graphical Forms and Operations
• Graphs as a Representational Form
– The standard for language representation
– Deals with Entities and Relations
– Quantitatively-based; visually manageable
• Graph-based Analytics
– Framework for Data Association as shown
– Evidential searching/matching (supports query-based,
discovery-based analysis)
• Variety of Graph-Matching paradigms, issues
– Stochastic due to tagged uncertainties in graph elements
– Incremental to handle streaming real-time data
– Large scale to handle “Big Data”; eg Cloud-based
Some further Impacts regarding Semantics
• Optimal strategies for semantic “control” –
control of semantic complexities
– Rigorous control of Ontologies
– Controlled vs Uncontrolled Languages
• Eg Battle Management Language
– Robust Text Extraction, NLP
– Role of Human Mediators in system architecture
• Speed (automation) vs semantic accuracy
• Semantic Uncertainty
• Vague predicates; issue of Truth—leads to 3-valued forms of
Uncertainty Representation
Some Impacts regarding
System Architectures and Analytical Frameworks
• Many problems are “Weak Knowledge” problems
wherein the extent of reliable a priori dynamic
knowledge about the domain is limited
• This motivates an approach that must combine
deductive and inductive (or abductive) methods in
an effective way
– These tend to require technologies that support discovery
and learning-based hypothesis-formulation strategies
• Methods such as Complex Event Processing,
Probabilistic Argumentation, Graph-based Relational
Learning are some of the new inferencing methods
being studied.
Representative Architectures:
Inductive + Deductive
Earliest Thoughts on Combining Inductive and Deductive Inferencing for Fusion*
* Integrating the Data Fusion and Data Mining Processes Ed Waltz, Natl Symp on Sensor and Data
Fusion, 2004
Representative Architectures:
Hard and Soft Fusion Processes; Disparate Analytic Tools
Intel Cell – or –Company Opns Intell Support Team
Evidence and Entity -estimate
Foraging Services
Sensemaking
Services
Analytic Support
Services
Enterprise Service Bus
on
on
g
-
Hard (sensor) fusion
on
n
g
encin
g
encin
Soft (intel) fusion
Information
(Evidence)
Services
(Sensor) Data and
Computational
Services
Core
Enterprise
Servces
Summary
• Requirements for Data and Information Fusion Processes
and Systems have gone far beyond the goal of estimating
properties and geometries of entities
– Dealing with complex Semantics, inter-entity Relations, Social
Media and other Contextual effects, complex Temporal
dynamics, and Heterogeneous Data have made the design of
IF Systems a markedly new challenge.
• Incremental advances and accomplishments are being
realized but there is much to be done
• Major advances are needed in dealing with more complex
inferencing challenges to support efficient learning and
discovery processes.
• New partnerships are needed across various
multidisciplinary areas in order to address these new
complexities