© 2017 BAE SystemsApproved for Pubic Release; Distribution unlimited
1
Machine Learning & Artificial Intelligence: Transforming Defense, Intelligence, & Security
Dr. Denis GaragićChief Scientist – Machine LearningSeptember 27, 2017
Credit: agsandrew / Fotolia
© 2017 BAE SystemsApproved for Pubic Release; Distribution unlimited
2
• Human-centric Injection of AI to Make Machines Smart Enough to Help Their Human Users
• Efficient Learning Machines – 1st of 5 Third Offset Pillars
• Find Patterns
• Represent Domains & Problems
• React When Human Reaction Time Too Slow
• Multiple Communities of Interest Have AI Requirements: Advanced Electronics; Autonomy; Cyber; Electronic Warfare (EW); Sensors; Command, Control, Communications, Computers, and Intelligence (C4I); Space
Advances in AI and Autonomy Lead to New Era of Human-Machine Collaboration and Combat Teaming
Credit: aroma360.com
© 2017 BAE SystemsApproved for Pubic Release; Distribution unlimited
• Patterns Are Ubiquitous, Yet Often Obscure or Hidden –Finding & Exploiting Them Yields Critical Advantages in Many Situations
• Cognitive & Adaptive Systems Provide Such Advantages
• Incremental, Multi-modality, Multi-Domain Data-Driven Learning Approaches Are Required
• Continuous Learning During Execution Applies Previous Knowledge to Novel Situations
How Do We Do It?
3
© 2017 BAE SystemsApproved for Pubic Release; Distribution unlimited
4
Solving Tomorrow’s Defense, Intelligence, & Security Problems Requires Fundamental Machine Learning Theory Development Using New Paradigms
© 2017 BAE SystemsApproved for Pubic Release; Distribution unlimited
5
• Real-Time Performance Analysis, Anomaly Detection, & Behavior Prediction
• Autonomy & Resilience for Uninhabited Vehicle Perception & Control
• Adaptive Cyber Security & Network Intrusion Detections Detection, Recognition, & Response
Activity-Based Intelligence AnswersDifficult Analytic Questions
Normalcy Learning
Behavior Prediction
Anomaly Detection
Track Data
© 2017 BAE SystemsApproved for Pubic Release; Distribution unlimited
6
• Finding Slow Moving Targets Under Foliage is Hard, Monitoring Their Behavior is Many-fold More Challenging
• Consistent Observation of Individual Movers – Human or Otherwise – is Not Realistic
• Accurate Estimation of Type & Intent of Current Activity Behavior of Moving Groups Supports Additional Objectives
Classifying Activity of Humans from Noisy, Sparse Radar Observations Becomes Possible
Human and Pack Animals
Small Boats
Humans Patrolling
Animals Grazing
© 2017 BAE SystemsApproved for Pubic Release; Distribution unlimited
Probabilistic Inference Automatically Aligns Geospatial Vehicle Motion Data
7
Process
Process
Heatmap of Raw MOVINT Overlaid on Google Earth
Aligned Results
Uncover Parking Lot
Activity Hidden in Raw Data
Access Points &
Drive-Thru Lane
Emerge from Noisy
Data
MOVINT Heatmaps(Sub-regions)
Strip Mall (Starbucks, gym, etc.)
McDonald’s
Raw MOVINT Accuracy Insufficient for Location-Based Analytics Requiring
High Precision / Recall Geospatial Queries
© 2017 BAE SystemsApproved for Pubic Release; Distribution unlimited
Data
2 GPixels
Intelligence
100K TracksReasoning by Bayesian Machine Learning
Classification TrackingPreprocessing Segmentation
Registered
Data
Target
Objects
Classified
TargetsTracks
8
Adaptive Generative Machine Learning Enables Applications Requiring Both Proactive (Online) & Forensic (or Batch) Data Analysis
© 2017 BAE SystemsApproved for Pubic Release; Distribution unlimited
9
• Learns / Adapts Under Uncertainty from Limited Multi-Modality & Multi-View Observations
• Learns Compact Representations of Objects of Interest Online
• Performs Accurate Object Recognition Within Small Onboard Size, Weight, & Power (SWaP) Envelope
Adaptive Reasoning Recognizes Known & Unknown Objects from Multi-Modality Data
© 2017 BAE SystemsApproved for Pubic Release; Distribution unlimited
10
• Automated Generation of Relevant Reports from Time-Varying Multi-Modality (e.g., Visual, Radar, Signals) Scenes
• Integration of Learning & Language Understanding Reasons Over Activities (e.g., Space & Time Relationships; Object Interactions; Other Contextual Information) in Scenes
• Multi-Object Detection & Tracking, Encoder & Decoder Modules to Process Imagery & Generate Descriptions
Combination of Probabilistic & Deep Learning Approaches Performs Actionable Intelligence Discovery & Exploitation from Multi-Sensor Streaming Data
© 2017 BAE SystemsApproved for Pubic Release; Distribution unlimited
11
• Rapidly Characterize Emerging Agile Threats – Radars, Radio Communication Devices, Interference (Jammers)
• Synthesize & Optimize Electronic Countermeasures
• Assess Response Effectiveness
Closed-Loop Rapid Learning and Reaction Capabilities Intelligently Counter Threats
Detection & Characterization
Response Optimization
Effectiveness Assessment
Radar / Radio Threats
© 2017 BAE SystemsApproved for Pubic Release; Distribution unlimited
Machine Learning Predictions Make Otherwise Unobservable Aspectsof Economic Threats to National Security Observable
12
Ground TruthNews Media Reports
Militants kidnap 4 foreign workers
MEND rebels attack an oil barge & seize 9
hostages Baker Hughes’ executive killed
Norwegian rig offshore Nigeria attacked & 16 crew members kidnapped
Companies Start Exit Nigeria Due to Delta
Kidnapping Chaos
1) MEND killed 10 Nigerian soldiers2) MEND abducted 7 foreign workers3) Nigerian soldiers attacked militant camp, in ensuing battle 9 Nigerian soldiers killed
MEND Set Off 2 Bombs in Southern
Nigeria
Examine Causes for Predicted Changes –Missing Ground Truth (Reported News)
2
20
6
10
14
18
16
12
8
4
0
Ru
n L
en
gth
(T
rad
ing
Da
ys)
Significant Historical Events (Nigeria 2006) Median of Run Length Predictive Distribution
MEND attack on 4/20
Jan Apr Jul JanOct
© 2017 BAE SystemsIntended to be: Approved for Pubic Release; Distribution unlimited
13
• Covert / Discreet Communication in Dangerous Environments
• Communication in Super-Noisy Environments
• Communication for Speech-Disordered Individuals
Non-Acoustic Speech CommunicationPossible via Mouthed-Speech Understanding & Transcription
-+
© 2017 BAE SystemsIntended to be: Approved for Pubic Release; Distribution unlimited
Defense, Intelligence, & Security Present Broad Range of Challengesfor Machine Learning & AI
14
Extremely Low Size Weight & Power Devices
Too Much Data Too Little Data
Massive Data Centers
Exquisite Data Noisy Late Data
Minutes / Hours / Days for Human Decisions Sub-Second Decisions by Machines
© 2017 BAE SystemsApproved for Pubic Release; Distribution unlimited
15
Thank you
BAE Systems – Technology Solutions
Dr. Denis Garagić
Burlington, MA, USA